The drugs are semaglutide, which Novo Nordisk sells as Ozempic and Wegovy, and tirzepatide, which Eli Lilly sells as Mounjaro and Zepbound.

Tirzepatide became the best-selling pharmaceutical in the world. Though AI revenue is accelerating and the gap is closing, through 2025, two molecules out-earned the two leading frontier AI companies combined, and they did it while reaching fewer than 2% of the roughly 800 million people who would qualify for them globally.

Despite mainstream adoption and public testimonials from celebrities like Oprah, Serena Williams, and Elon Musk, in America alone, under 6M of the 137M clinically eligible adults are treated. Goldman Sachs and Morgan Stanley both project 15 to 20% penetration by the early 2030s, about 3-4x what it is now.

The Indication Cascade

A hormone is a chemical messenger that travels through the bloodstream, meaning it reaches every tissue the blood reaches. Nothing about the molecule changes between those tissues, only the cell receiving it.

Testosterone is the clearest example. The same molecule tells hair follicles on the scalp to shrink, follicles on the face to grow, raises competitive drive in the brain, and stimulates sperm production in the testes.

GLP-1 works the same way. Its receptors sit on the insulin-producing cells of the pancreas, on neurons in the hypothalamus, on the stomach lining, and in the brain’s reward circuitry. This is why we see such varied changes across the whole body.

Every GLP-1 approval since 2017 has surfaced an effect nobody designed for, and a later trial has turned that effect into the next approved indication. Diabetes patients lost weight, which produced the obesity label in 2021. The obesity drug then cut major cardiovascular events by 20%, which produced the cardiovascular label. A kidney trial cut major kidney outcomes by 24%, and on the day those results were announced, shares in DaVita and Fresenius, the two largest dialysis operators in America, each fell about 17%, because a drug that slows kidney decline shrinks the dialysis population. The pattern has since run through sleep apnea and liver disease.

Each turn of that cycle adds tens of millions of newly eligible patients and reprices an industry that was not in the conversation the year before. Walmart’s U.S. CEO said the company could already see GLP-1 users buying smaller baskets, and Nestlé now sells a frozen-meal line designed for them.

So how far does the cascade run? In January 2025, Washington University followed 215,970 patients on these drugs, matched to controls, across 175 conditions. Risk fell on 42 of them and rose on 19.

Of the 42, the one with no incumbent is addiction. GLP-1 receptors sit in the brain’s reward circuitry, and the same mechanism that quiets food noise appears to dampen the response to a drink or a drug. Around 36 million Americans are affected by alcohol, opioids, and stimulants.

This leads to an ever-growing addressable market for these companies as new trials reveal further effects.

My researchers at Social Capital spent months on this topic and created a 99-page deep dive that starts with why the body defends the weight it already has and ends with why this drug matters beyond obesity.

One of the many points we unpacked: GLP-1 is the first commercially significant proof of concept. Find the hormone the body uses to run a system, engineer a version that survives in the bloodstream, and you treat the system instead of its symptoms.

Here’s what else you can expect to learn:

• Why no third company has built one of these, and what a rival would have to assemble to try.

• The receptor the industry wrote off, how Lilly turned it into the best-selling drug on earth, and the scaling rule every generation since has followed.

• How a hormone that survives two minutes in the bloodstream became a once-weekly injection, and the desert lizard that made it possible.

• All 175 outcomes measured, the 19 that got worse, and what the confirmatory addiction trials could do to the size of this market.

• What stops this: who quits, what comes back when they do, and the four places the pipeline is looking for the fix.

If you’re interested, learn with me below. Hope you enjoy reading!

Chamath

Disclaimer: The views and opinions expressed above are current as of the date of this document and are subject to change without notice. Materials referenced above will be provided for educational purposes only. None of the above will include investment advice, a recommendation or an offer to sell, or a solicitation of an offer to buy, any securities or investment products.

Deep Dive PDF below ↓

Read more

1) The Leverage Inside Korea’s Memory Trade

South Korea’s KOSPI, its benchmark stock market index, has triggered consecutive circuit-breaker sessions this week. At the low, the index had erased almost 40% of its value from the record above 9,380 it set little more than a month earlier. As much as $2.18T in Seoul market value had disappeared.

Korea had listed single-stock leveraged ETFs nine weeks earlier, on May 27. These funds promise twice the daily move of one company’s shares, and retail money crowded into the two tracking Samsung and SK Hynix.

Both were hit hard on July 29: SK Hynix fell nearly 20% before closing down 9.6%, while Samsung dropped as much as 14% before ending 5.2% lower. Because the two make up more than half of the KOSPI, their losses drove the index sharply lower. To maintain twice the daily exposure, these funds generally add exposure after gains and cut it after losses, creating further volatility and cascading risk.

In addition to the broader reassessment of the AI trade this month, CXMT, China’s largest DRAM producer, rose 466% in its July 27 Shanghai debut. It became the mainland’s most valuable listed company. Although CXMT still trails Korean producers in scale and advanced memory technology, the listing raised concerns that rising Chinese capacity could pressure conventional memory prices. Zooming back out, the KOSPI is still up over 50% YTD, the strongest performance among major markets.

On July 31st, the Minister of Finance and Economy, Koo Yun-Cheol, announced South Korea’s plan to launch a new sovereign wealth fund next year, specifically for long-term investments in strategic domestic industries, including AI, chips, robotics, defense, and biotechnology. Seoul will put 20 trillion won (~$14 billion) into the fund, with 16 trillion won in capital from shares it holds in state-run financial institutions and 4 trillion won from inheritance and gift taxes.

2) Inside the Fight for AI Governance

On July 27, Nvidia and dozens of partners launched the Open Secure AI Alliance, six days after OpenAI disclosed that an internal cyber-evaluation agent escaped containment and breached Hugging Face.

Closed AI tools blocked parts of Hugging Face’s forensic work, while a self-hosted open-weight model, GLM 5.2, reviewed more than 17,000 actions and helped contain the intrusion. Open weights make a model’s trained parameters publicly available, allowing organizations to run and inspect them on their own infrastructure. Hugging Face joined the alliance while leading closed frontier labs OpenAI, Anthropic, and Google did not.

Three other documents set out competing positions within the same timeframe. Nvidia’s July 24 letter, Open Weights and American AI Leadership, began with 25 signatories and soon climbed to over 230. Anthropic remains absent and, on July 27, responded that it opposes a categorical ban on open weights but supports controls on advanced chips, limits on industrial-scale model copying, and mandatory safety testing for sufficiently capable models.

On July 28, more than 1,100 industry employees signed a letter titled Pacing the Frontier, asking Washington to build international mechanisms to slow the development of automated AI. Anthropic CEO Dario Amodei and OpenAI chief scientist Jakub Pachocki signed the statement, and both companies endorsed it.

These documents focus on how capable models should be governed and deployed, with a recent emphasis on open-weight models. Beginning in August, a June 2 executive order directs Treasury, the Department of War through NSA, and DHS through CISA to develop a classified benchmarking process and a voluntary pre-release framework for covered frontier models.

3) Citadel Steps In as Situational Awareness Face Fire Sale

Leopold Aschenbrenner launched Situational Awareness in late 2024 with several hundred million dollars of seed capital. He named it after his essay on artificial general intelligence and ran it as a concentrated long bet on AI equities.

The fund returned 439% after fees in the first half of 2026, while reported assets peaked at $45B. The fund reportedly carried gross exposure of up to four times its capital. At that level, $100 of investor equity supports $400 of positions. A 10% move against the portfolio would erase about $40, or 40% of the fund’s capital. As losses reduce that cushion, prime brokers can demand more collateral or force positions to be cut.

After this month’s AI sell-off, on July 30th, Leopold sold most of its public-equity portfolio to Citadel. Situational Awareness still holds around $10B worth of assets, and the largest of them remain in private companies like Anthropic.

Learn With My Friends and Me…

Chip Stocks Crash, $20B Fund Margin Called, Frontier Labs: SLOW DOWN AI, Mamdani’s Grocery Stores

Other Reading…

• 22580: From GPT2 to Kimi3, Explained (@waterloo_intern)

• Why Compute Might Get 10x+ More Expensive in Coming Years (Dwarkesh Patel)

• The Actual Reason Why Google “Fell Out” of the AI Race Changes Everything (Alberto Romero)

On X…

In 2025, there were over 36K traffic fatalities on US roads. Waymo’s peer-reviewed safety data across 170 million driverless miles shows 92% fewer serious crashes or fatal injuries than human drivers in the same conditions. The autonomous driving revolution is removing one of the largest causes of accidental death in America.

2025 was the inflection point:

• Autonomous ride-hailing is now commercially available in 28 cities across five countries

• Waymo surpassed 20 million cumulative paid driverless rides by the end of 2025 and is now delivering 500,000 rides every week

• Last-mile delivery robots completed millions of autonomous deliveries on college campuses and city sidewalks across the US

The key metric to watch will be the price of a mile.

A personally owned vehicle costs $1.10 per mile when you factor in insurance, maintenance, depreciation, and fuel. The price gap between autonomous cars and personally owned cars continues to close as fleets scale and hardware costs fall. ARK Invest estimates the robotaxi market at $11 trillion once cost per mile hits $0.25.

I’m keeping a close eye on this sector:

And robotaxis are only one piece.

Goldman Sachs expects 65% of US car sales could be autonomous vehicles by 2040. Aurora plans 200+ driverless trucks by the end of 2026 and thousands by the end of 2027. Autonomous trucks target the $900 billion US trucking industry, where labor accounts for 45% of per-mile costs. Last-mile delivery robots are also solving the most expensive leg of e-commerce logistics. Additionally, the industries that exist because of human error, including car insurance, personal injury law, and traffic enforcement, face a structural rethink.

My research team at Social Capital mapped the technology, the economics, and who actually wins. The result is a 92+ page deep dive.

Here is what we cover inside:

1. What happens inside an autonomous car and why it took 30 years to get here.

2. The six levels of autonomous driving, where we are today, and what Level 5 actually requires.

3. Tesla vs. Waymo: two completely different bets on the same problem and their software strategy

4. How safe and cheap autonomous driving already is compared to a human driver, mile for mile.

5. The $11 trillion robotaxi race: who is winning, where regulation is the real bottleneck, and what the Arizona model shows.

6. How China is closing the gap on Waymo in 2026

7. The industries beyond ride-hailing that get disrupted when human error disappears

Read the deep dive and let me know what you think.

Chamath

Disclaimer: The views and opinions expressed above are current as of the date of this document and are subject to change without notice. Materials referenced above will be provided for educational purposes only. None of the above will include investment advice, a recommendation or an offer to sell, or a solicitation of an offer to buy, any securities or investment products.

Deep Dive PDF below ↓

Read more

How to think about the distillation and banning open source debate?

Distillation means running someone else’s model at scale: you fire up someone else’s model, ask it questions, watch the answers, and use those answers to train your own model. Do that tens of millions of times and you’ve exfiltrated trillions of question-and-answer pairs.

The labs frame this as industrial-scale theft that warrants government protection.

I think it is a red herring.

If you actually wanted to stop it, you’d KYC your customers. Force a real identity, put them behind a bounded credit card, and the industrial-scale account farms die overnight. It would slow your revenue, so the labs are not doing so. Instead, they are asking Washington to ban the competition.

And everyone has distilled from everyone. Anthropic trained on the publishers, then paid a $1.5B fine for it. The Chinese labs distilled the Americans.

So why the sudden panic?

Because the model layer is commoditizing faster than anyone expected. The day you publish your benchmark scores, someone matches them within weeks, while the closed labs are still priced at 25 to 50 times the open alternative. A lot of what you’re watching is valuation preservation.

The current moat is above the model and below it.

Up the stack, in the applications people actually pay for, and down the stack, in the infrastructure, the chips and the cloud. (Read our AI Stack deep dive to understand the layers fully.)

I think we should not defend a duopoly in D.C. and go win the layers that are defensible. If the government steps in to save them, it will just tax every American company that buys AI, and the market will decapitate the trade.

Think and then answer this question for yourself.

Caught My Eye…

1) When AI Solves the Unsolved and Deceives Its Makers

A conjecture is a claim mathematicians believe but cannot prove, sometimes for generations. This month, a multitude of them, some open for 40 to 90 years, fell in days.

On July 19, Anthropic mathematician Levent Alpöge announced that Claude Fable 5 found a counterexample to the Jacobian conjecture, open since 1939. Within days, Terence Tao had worked through why it holds, and the example was machine-checked in formal proof software.

The result followed OpenAI’s May announcement that an internal general-purpose model had disproved Erdős’s conjecture, open since 1946. Several other AI-assisted counterexamples have appeared, although they vary considerably in importance and verification status.

Mathematics gives AI unusually clear feedback because many proposed answers can be checked through calculation, expert review, or formal proof software. These results provide strong evidence that frontier models can contribute original mathematical results, especially when a large search space is paired with an objective way to score candidates. Similar generate-and-verify systems may eventually prove useful in fields such as algorithm design, chip engineering, materials science, and drug discovery, where candidate solutions can be tested against clear constraints.

The same kind of model has another side. On July 20, OpenAI disclosed that the system it credited with disproving the Erdős conjecture had repeatedly worked around its own controls in testing.

It ignored directions to report only in Slack, and instead worked to find a flaw in its sandbox to open a public code request. Many safety controls for AI assistants are designed around individual actions. If an action is disallowed, it is blocked, or the model must ask for explicit approval. But in long-running models capable of working on extended tasks, there appear to be new behaviors that circumvent these rules by learning the blind spots of the approval system to achieve its goals. For example, a model was able to split an authentication token into fragments and slip past a scanner, making each individual action look acceptable, but creating an outcome that was not approved.

The same week, the UK AI Security Institute reported that every frontier model it tested tried to cheat on its evaluations. Models did not reliably report this behavior when asked, and often did not reason about it in their chain-of-thought, suggesting that detecting cheating will likely require robust monitoring methods.

2) Travis Kalanick’s $1.7B Bet on Physical AI

On July 22, Travis Kalanick announced a $1.7B raise for Atoms, led by Andreessen Horowitz, with Ben Horowitz taking a board seat.

Kalanick has spent his career applying software to physical-world industries. Uber built a digital network for moving people. CloudKitchens applied a similar model to food production, treating commercial kitchens like computing infrastructure. The kitchens acted as the processors, converting ingredients into meals, while the real estate provided the physical capacity needed to operate and scale them.

Atoms expands that idea across the industrial economy. Kalanick asks: “What about an OEM that builds atoms-based computers for all the major industrial sectors?”

The company is betting on Industrial AI: systems that combine software, sensors, robotics, and AI to automate how physical goods are made and moved. Atoms brings together CloudKitchens and its food-robotics business, a mining unit built on industrial automation company Pronto, and a self-driving freight effort.

The pitch is that the world of atoms is at the cusp of a new industrial revolution, where what happened to the digital world of bits can now be applied to the physical world, unlocking trillions of dollars worth of productivity.

3) Google’s Biggest Quarter Ever

On July 22, Alphabet reported the largest quarterly profit in its history. Net income reached $112.1B, up 298%, or $9.11 per diluted share, on revenue of $119.8B. The result included a $99B net gain on Alphabet’s equity holdings, which generated $98B of net other income. Alphabet said the gain added $6.26 to earnings per share, implying EPS of roughly $2.85 without it, slightly below the $2.88 to $2.89 analysts expected.

Alphabet said the gain came primarily from SpaceX and an unnamed private company. Anthropic is a likely contributor: Alphabet reportedly owns about 14% of the lab, whose valuation rose from $380B to $965B after a $65B funding round during the quarter.

Operating cash flow was $39.1B, while capital spending roughly doubled to $44.9B, resulting in free cash flow of negative $5.9B, its first negative quarter according to Reuters.

Alphabet also raised its 2026 capex guidance to $195B to $205B, from $180B to $190B. The clearest operating strength was Google Cloud, where revenue rose 82% to $24.8B and operating income reached $8.8B, producing a 35.6% margin.

Learn With My Friends and Me…

The Fight Over Open Source AI, Anthropic’s $1.5B Payout, NYC Socialists: Evictions = Violence?

Other Reading…

• Why We’re Buzzing (Jack Dorsey)

• Why Hasn’t AI Increased Unemployment? (Peter McCrory)

• Agentic AI - The Killer Use Case for Blockchain and Crypto (Franklin Templeton)

• Creativity, Intelligence, and What We’re Building Toward (Berggruen Institute)

On X…

Governments are using these same capabilities. Whether the target is a company or a country, attackers usually get in through a person or a technical weakness. That can mean a phishing email, stolen password, unpatched device, misconfigured system, or newly discovered software vulnerability. From there, they try to become an administrator, blend into normal activity, and reach systems they can steal from or disrupt.

This creates a fundamental imbalance. The attacker only needs one path that works. The defender has to protect every device, account, employee, vendor, and piece of software, every day. That is why a cheap phishing email can sometimes reach the same target as a multimillion-dollar software exploit.

AI is now changing the speed and cost of each step. It can research targets, write personalized messages, generate exploit code, test vulnerabilities, steal credentials, and move through networks much faster. In one operation disclosed in 2025, an AI handled 80 to 90% of the tactical work on its own.

In OpenAI’s test, the models were supposed to operate inside an isolated environment with no direct internet access. To solve a cybersecurity benchmark, they found a previously unknown flaw in the software controlling that environment, used it to reach the open internet, and then chained additional vulnerabilities and stolen credentials to run code on Hugging Face’s servers. From there, they accessed the benchmark answers stored in Hugging Face’s production database, effectively cheating on the test.

In OpenAI’s evaluation, the result was a higher benchmark score. At a company, the same attack path could reach customer data, payments, or core operations. At a national level, it could reach power grids, communications networks, ports, and financial systems.

China-linked operators have maintained access inside parts of US critical infrastructure for at least five years. North Korea has stolen $6.75 billion in cryptocurrency. Russia’s NotPetya attack caused more than $10 billion in global damage.

Taken together, these examples show how a small number of operators can threaten systems that took decades and billions of dollars to build. AI can increase that leverage by reducing the time and expertise required to run an attack.

An understanding of cyberwarfare now helps you separate real capability from headlines, assess where companies and governments are most exposed, and understand how AI changes the balance between attackers and defenders.

That is why our research team put together a cyberwarfare deep dive. Inside, we explain:

• Why China has spent years inside the US systems that would matter most in the first 72 hours of a Pacific conflict

• Why a $50 phishing email can sometimes reach the same target as a $5 million zero-day, and why that math favors smaller states

• What changes when AI can run 80 to 90% of an attack, and why human-speed defense may no longer be enough

• Why no leader publicly threatens a cyberweapon, and why revealing one can destroy its value

Hope you enjoy reading (and stay safe out there!)

Chamath

Disclaimer: The views and opinions expressed above are current as of the date of this document and are subject to change without notice. Materials referenced above will be provided for educational purposes only. None of the above will include investment advice, a recommendation or an offer to sell, or a solicitation of an offer to buy, any securities or investment products.

Deep Dive PDF below ↓

Read more

This week, I want to share two: one I have discussed recently, and another that prompted some interesting responses from others (my What I Read This Week is still below).

If you find this useful, let me know in the comments, and we will do more every week.

As always, don’t just take my answers or anyone else’s. Think and then answer these questions for yourself.

Share your answer in the comments; we will repost the sharpest take.

Question 1) With massive AI bills hitting enterprises, how should companies think through maximizing ROI for their token costs?

Most CEOs and CFOs have no idea how much tokenmaxxing is happening within their own companies, and it will show up in the numbers.

At 8090, my CTO told me our token costs are doubling roughly every 45 days, and the incremental productivity we get from that doubling is maybe 5-10% at most.

Now run that against a Fortune 500 company that said it is all in on AI. The spend can quietly compound inside OpEx, and suddenly a quarter gets missed by a few pennies of EPS. The CEO would ask the CFO where the money went, and they would likely trace it back to the $56-per-million tokens of intelligence, when a $0.50 version could have done the job.

So how do you not become that company?

First, stop paying for the flagship where you don’t need it. The cheap models are now 80% to 95% as good as the frontier models on most tasks. If Pepsi is a fraction of the price of Coke and tastes almost the same in the dark, you at least have to manage that as a risk. Pick the model per task, route it through a control plane that gives you optionality, and pay frontier prices only for the few jobs that genuinely need frontier intelligence.

Second, watch where your data goes. When you pipe your proprietary workflows and your competitive logic through a closed frontier model, you are renting intelligence and giving away your edge to providers that may also want to compete with you.

Measure the output you get, and treat intelligence like any other input cost.

Question 2) If every smart enterprise starts to pull its AI in-house to protect its edge, who’s left to pay for the $1.4T buildout?

Gavin Baker’s answer:

Replit CEO Amjad Masad’s answer:

Caught My Eye…

1) Open-Weight AI Reaches the Coding Frontier

On July 16, Chinese lab Moonshot AI released Kimi K3, a 2.8-trillion-parameter model that took first place on the Frontend Code Arena, a leaderboard scored by blind developer votes on the web apps models generate. K3 reached 1,679 points, ahead of the closed flagships Claude Fable 5 and GPT-5.6 Sol, and led six of the seven frontend categories.

On broader intelligence tests it still trails both, but at $3 and $15 per million input and output tokens it undercuts Fable 5’s $10 and $50, and Moonshot plans to release the full weights by July 27, which would make it the largest open-weight model yet. Until then, every K3 spec is Moonshot’s own claim.

The same push is coming from US startups selling the means to build and own AI rather than rent it.

On July 15, Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, released Inkling, its first model, with open weights under an Apache license that lets anyone download and modify it.

Inkling is a 975-billion-parameter mixture-of-experts model that activates only about 41 billion parameters per request, which makes it cheaper to run. Thinking Machines describes Inkling as a base for customization rather than the strongest model available, and its full weights are free to download. The company’s paid product is Tinker, a fine-tuning tool that adapts models to a customer’s data. Bridgewater’s AIA Labs published a study that used Tinker to fine-tune an open model for financial document triage. The custom version scored 84.7% accuracy, compared to 78.2% for the best frontier model tested, at roughly one-fourteenth of the cost per task.

A week earlier, on July 8, Prime Intellect raised a $130 million Series A at a $1 billion valuation. Prime Intellect rents decentralized computing and open-source software that let a company train its own agents, reporting a $100 million annualized revenue run rate. A case study with Ramp showed their post-trained model beating Opus at spreadsheet search, and running 27% faster and far cheaper than Haiku.

2) Ex-SpaceX Engineers Raise $115M to Revolutionize Construction

On July 14, TerraFirma raised about $115 million, including a $100 million Series A led by Kleiner Perkins, and said it will hire 300 people and build a Texas factory. It was founded in 2024 by former SpaceX engineers Noah Schochet and Noah McGuinness, who met at Princeton and worked on Starlink and Starship.

The company retrofits heavy construction machinery, such as excavators, dozers, and loaders, into robots that a skilled operator operates from a screen rather than climbing into the machine. Using an interface it calls “click to dig”, an operator sketches the required work as a 3D model in about 60 seconds, and the machine runs the task for more than 20 minutes on its own, with an Xbox controller kept as a manual backup. One operator can currently handle three to five machines, which TerraFirma says makes each operator up to 300% more effective and the work safer, since no one has to sit in the machine.

US construction labor productivity has fallen about 0.6% every year since 1965 even as productivity across the broader economy has grown at roughly 1.6% annually. It’s estimated that the industry needs 349,000 more workers in 2026 just to keep pace. TerraFirma runs its own projects rather than selling the technology, which gives it live job sites and field data. The field is heating up as Caterpillar and Komatsu already offer remote control, and Bedrock Robotics raised $270 million to field operatorless excavators this year.

Learn With My Friends and Me…

ALL-IN: Can the AI Industry Regulate Itself? Stripe Wants PayPal, China Catches Up, NY Bans Datacenters

Other Reading…

• Making Fable Cheaper Than Opus (Joon Lee)

• You Just Hired a Million Bad Employees (George Sivulka)

• A Framework for Frontier AI and the Dawning of a New Age (Demis Hassabis)

• The Reverse Information Paradox (Satya Nadella)

On X…

I want to further make our research and weekly publications even more useful to you. As a member of our community, could you take just 2 minutes to complete this survey:

Start Survey

There is one free-form Q&A question at the end. I’d love to answer a few from everyone this week.

This will help us tremendously in…

Read more

Caught My Eye…

1) The New Models: Grok 4.5, GPT-5.6, and Muse Spark 1.1

Three of the largest AI builders shipped new models within a day of each other. SpaceXAI released Grok 4.5 on the 8th. The next morning OpenAI began rolling out GPT-5.6 as three models named Sol, Terra, and Luna, and Meta launched Muse Spark 1.1. Meta made the largest structural shift, moving from open-weight model releases to their first paid closed model.

Each of them pushed for more affordable models while still maintaining performance.

• Grok 4.5 costs $2 per million input tokens and $6 per million output tokens.

• OpenAI’s smallest new model, Luna, lands at $1 and $6.

• Meta’s Muse Spark 1.1 runs at $1.25 and $4.25.

Set that against the flagships still on sale: Anthropic’s Opus 4.8 at $25 for output, OpenAI’s own Sol at $30, and Anthropic’s Fable 5 at $50.

They also converge on what the models are built to do: run tools and operate software, rather than only answer questions. Meta built Muse Spark 1.1 for agentic work: a 1-million-token context window, the ability to run subagents in parallel, and training to click through desktop, mobile, and browser interfaces on a user’s behalf. Grok’s selling point is efficiency, finishing a coding task in roughly a quarter of the tokens Opus 4.8 uses. GPT-5.6 was directly implemented into their new product ChatGPT Work, a way to integrate your entire workflow fully.

By the makers’ own numbers, each model leads on a different axis, with none fully sweeping all benchmarks. On the same day, Claude decided to reset all 5-hour and weekly rate limits for their users.

2) DeepSeek Designs Its Own Inference Chip

On July 7, Reuters reported that DeepSeek is designing its own AI chip, which would allow it to rely less on Nvidia and Huawei. DeepSeek is the Hangzhou-based company whose cheap, open models jolted the industry in early 2025, and it is now widely regarded as China’s national AI champion. The chip targets inference, the stage where a trained model answers a user, rather than training, the far more expensive stage where the model is built. The effort is about a year old and still early, with the company in talks with chip designers, foundries, and memory suppliers.

Designing the chip is only the first hurdle. U.S. export rules cover certain foreign-made products built with American software or equipment, requiring suppliers such as TSMC, Samsung, and SK Hynix to obtain U.S. approval before providing controlled technology to China. That limits Chinese companies’ access to both leading-edge chip manufacturing and the advanced high-bandwidth memory needed to keep AI processors running at full speed. A strong design on paper does not become a competitive chip without both.

DeepSeek would be competing at home with Huawei and other tech rivals, including Alibaba and Baidu, who are already developing their own AI chips. Alibaba has its Qwen models, Baidu has ERNIE, and even Zhipu AI, the creators behind GLM, is starting to explore its own custom AI chips.

3) Microsoft Cuts 4,800 Jobs and Sells Four Studios

On July 6, Microsoft cut about 4,800 jobs, representing 2.1% of its workforce. Currently, Microsoft is the worst-performing Magnificent 7 stock this year, down more than 18%.

The majority of the layoffs came from its Xbox division, and four gaming studios will be spun out to be independent again. Xbox CEO Asha Sharma told the team, “Our business today is not healthy.“ It is on track to end the fiscal year at about a 3% margin, which is 3-10x lower than its competitors, and stands in stark contrast to Microsoft’s roughly 39% company-wide net margin.

Gaming revenue fell about 6% over the nine months through March, and console hardware sales dropped 33% from a year earlier.

Two forces are pulling the margin down. Game Pass, the $20-a-month subscription that delivers new titles to players on release day, undercuts the $70 those players used to spend per game. By Sharma’s own account, the bets on Game Pass and a broader content portfolio did not grow at the pace they expected while the core business also weakened. Buyers aged 18 to 24 spent roughly 25% less on games than in 2024 while playing more hours, most of them inside free platforms like Roblox and Fortnite.

It was also revealed that Asha Sharma will be joining the new Fed task force on Productivity and Jobs, along with Marc Andreessen of A16Z, and Charles I. Jones, Stanford professor of Economics who is currently working at Anthropic. They are tasked with assessing the economic impact of new general-purpose technologies such as AI and helping inform the Federal Reserve’s policy judgments.

At the same time, news broke out that companies like Starbucks are now developing AI software that could replace several enterprise applications currently supplied by Microsoft.

Learn With My Friends and Me…

OpenAI vs Anthropic IPOs, Anthropic $3T, Zuck’s Price War, China Ends Open Source?, Trump Accounts

Other Reading…

• AI’s Value Capture Problem (Jaya Gupta)

• 46 Thoughts on the Near Future (bayeslord)

• Institutional Sovereignty in the Age of AI (Palantir)

On X…

Every transaction on the Bitcoin network is public. Instead of a bank or other trusted intermediary privately verifying payments, thousands of independent computers verify every transaction by checking the same public ledger.

That design decentralized verification, but it also made every coin and every wallet traceable. At the coin level, that visibility breaks one of the most important features of money: fungibility.

Something is fungible when its individual units are identical, indistinguishable, and interchangeable. It ensures that the exact origin, specific unit, or history of an item does not affect its value or utility.

Bitcoin’s public ledger lets anyone trace a coin’s history, including whether it has ever been linked to illicit activity. Exchanges often reject tainted coins or only accept them at a significant discount. Because a coin’s history affects its value, one bitcoin doesn’t necessarily have the same market value as another.

At the wallet level, Bitcoin’s transparency can expose individuals. A Bitcoin address is more like a pen name than a bank account. It remains pseudonymous only until someone links it to a real identity. One exchange signup, public payment, or leaked database can make that link. Once a wallet is linked to a real identity, every transaction ever sent from that address becomes readable, retroactively and forever.

To address this privacy issue, developers went back to a discovery made decades ago. In the 1980s, MIT researchers asked whether you could prove a statement is true while revealing anything else about it. Their answer: the zero-knowledge proof.

A simple analogy to help grasp how they work:

Suppose I claim two balls are different colors, one red and one green. You are colorblind, so they look identical to you.

You put both balls behind your back, secretly decide whether to swap them, then show them again. If the balls really are different colors, I can tell whether they changed hands. If they are identical and I am bluffing, I can only guess.

But after enough random rounds, repeated correct answers become nearly impossible unless the balls really are different colors. I prove I know whether the balls changed hands, but you never learn the hidden information: the balls’ colors.

The same logic applies to private digital money.

In a private transaction, the hidden information is the payment itself: who sent it, who received it, which coins moved, and how much was sent. A zero-knowledge proof turns that hidden information into a public receipt the network can check. If the sender owns the coins, has not already spent them, and has followed the rules, the receipt passes. If any of those facts are false, it fails.

The network verifies the payment without seeing the payment.

Governments remain deeply divided over what private digital money should look like. As rules evolved, 73 exchanges delisted privacy tokens in 2025, yet they became the best-performing sector in crypto.

My research team at Social Capital traced financial privacy from first principles, exploring the monetary history of Mesopotamian clay tokens to the advanced mathematical proofs live on-chain today. Inside the deep dive, you will find:

• The four fundamental properties of money explain why gold and cash can pass the fungibility test, whereas Bitcoin fails.

• The cryptography that lets a transaction prove itself while revealing nothing.

• A landscape of privacy tokens: Monero’s mandatory privacy, Zcash’s opposite bet, and the first private digital dollars.

• The compliance split that divides the sector, and who has built a way through it.

• The three paths this could resolve over the next few years.

• Which privacy token designs actually work, and which pose the biggest risks?

• The strategies governments are deploying to push or rein in privacy tokens.

Money that reveals nothing and money that reveals everything are being built at the same time. Read it and share your perspective in the group chat.

Chamath

Disclaimer: The views and opinions expressed above are current as of the date of this document and are subject to change without notice. Materials referenced above will be provided for educational purposes only. None of the above will include investment advice, a recommendation or an offer to sell, or a solicitation of an offer to buy, any securities or investment products.

Deep Dive PDF below ↓

Read more

Caught My Eye…

1) The Sovereign AI Trade

On June 29, Palantir and Nvidia announced a partnership to run AI inside classified and air-gapped government systems, meaning networks cut off from the public internet. The deal pairs Nvidia’s chips and open Nemotron models with Palantir’s data software, enabling agencies to run modern AI on secret data without it leaving their control. As these customized models are used in production, teams can continually improve them within their own environments using new data and feedback.

Three days later, Microsoft made the same case. They launched Microsoft Frontier Company, a new business with a $2.5 billion budget and 6,000 engineers who embed inside customers to build and run their AI. Microsoft called it a step “beyond forward-deployed engineering,” in which a customer’s data, IP, and expertise are never used to train models that would commoditize them.

“The last thing any of us want is a world where every company across every sector is ceding value to a few models that eat everything they see… There is no societal permission for an AI future that hollows out entire industries.” - Satya Nadella, CEO of Microsoft

There are also discussions about how the public should have a stake in that AI future. On July 2, the Financial Times reported that OpenAI has discussed giving the U.S. government a 5% stake, worth about $42.6 billion against the $852 billion valuation it set in March. The pitch is to establish a sovereign wealth fund structure in which Washington would hold a 5% stake in each major U.S. AI company. The idea is to give the public a financial interest in AI companies so they can share in the upside.

2) Meta Moves to Sell Its Excess AI Compute

On July 1, Bloomberg reported that Meta is developing plans to build out a cloud business, called Meta Compute, to sell its spare AI computing power and hosted models to outside customers. The effort is run by infrastructure chief Santosh Janardhan, Meta Superintelligence Labs’ Daniel Gross, and company president Dina Powell McCormick.

This idea follows up on what was discussed at their May shareholder meeting. Zuckerberg said, “It’s definitely on the table,” in response to a question about potentially competing with Amazon and Microsoft in cloud computing. He also discussed the opportunity last year, sharing that “every week there are different companies that come to us from outside asking us to both stand up an API service or asking if we have compute that they could buy from us at some premium to what we’ve bought it at.”

The move mirrors SpaceX and other capacity holders that turned spare infrastructure into a revenue line. If this initiative works out, Meta will join Amazon, Microsoft, Google, and Oracle in the public cloud arena, with only Apple remaining the sole private compute hyperscaler.

3) Aramco Backs $8.3B Open-Model Neocloud

On July 1, Together AI raised $800 million at an $8.3 billion valuation, led by Aramco Ventures, with Nvidia and others joining. Together AI is a neocloud that rents out GPU computing power but bundles that hardware with its own software for running open models like DeepSeek and Kimi. It says the combination reduces the cost of serving a model by up to 80% compared with closed systems from OpenAI or Anthropic.

Annual bookings crossed $1.15 billion last quarter, with customers including Cursor, Cognition, and Decagon, and open-model use across the industry tripled in twelve months. Together has lined up commitments for more than 500 megawatts of compute, with expectations to grow capacity roughly 50-fold over five years.

The same day, Venice AI raised $65 million at a $1 billion valuation, with crypto fund Dragonfly leading the round. Venice is a privacy-first alternative from ShapeShift founder Erik Voorhees that encrypts prompts on your device and keeps no record of your chats. It is already profitable on about $70 million of annualized revenue with 3 million users.

Learn With My Friends and Me…

All-In: I Sovereignty Wars, Palantir-Nvidia Deal, SCOTUS Birthright Ruling, Newsom’s CA Budget Lie

Other Reading…

• Our Thoughts on the Importance of AI Sovereignty (Palantir)

• A Frontier Without an Ecosystem is Not Stable (Satya Nadella)

• Nvidia, CoreWeave, and Nebius: Inside the Circular Financing of the GPU Boom (IO Fund)

• AI Company Anthropic Announces It Will Begin Developing Drugs of Its Own (STAT News)

On X…

China has led the trend for years, reducing its U.S. Treasury holdings while buying gold. Its gold reserves now exceed 74m ounces, according to its central bank.

Gold has also become a yardstick for the rest of the market. In dollar terms, the U.S. stock market continues to set record highs. But measured in gold rather than dollars, the S&P 500 has not made a new high since 2000. Measuring financial assets against gold helps strip out the effects of monetary expansion because policymakers cannot create more gold.

Gold’s rise has also been matched by the other precious metals. The last comparable cycle came in the late 1970s, a decade that ended with 20% interest rates and a monetary policy reset.

But the forces moving each metal are different.

Gold is being treated as a hedge against uncertainty. Silver spiked above $100 per ounce in January, sending manufacturers racing to redesign their products around a cheaper metal. Platinum and palladium are shaped more by industrial demand, emissions rules, and the shift to electric vehicles.

But those forces only explain part of the story. Each metal has a different constraint that determines how far prices can move and what happens next. That is why we created this primer to map what drives each one and what its price may be signaling.

Inside our 82-page primer, you will find:

• What function do Precious metals serve in the modern global system

• How do supply, scarcity, and demand regimes differ across metals

• How are Precious metals interpreted in portfolio construction and risk frameworks

• Which central banks are buying gold, what they are selling, and why

• Why silver prices barely influence how much silver gets mined

• How gold-backed digital tokens and BRICS fit into the dollar system

If you want to learn with me, read the deep dive below and tell me what you think.

Chamath

Disclaimer: The views and opinions expressed above are current as of the date of this document and are subject to change without notice. Materials referenced above will be provided for educational purposes only. None of the above will include investment advice, a recommendation or an offer to sell, or a solicitation of an offer to buy, any securities or investment products.

Deep Dive PDF below ↓

Read more

Caught My Eye…

1) Training Physical AI on Video Game Clips

On June 25, General Intuition raised $320 million at a $2.3 billion valuation, led by Khosla Ventures, with General Catalyst, Jeff Bezos, and Eric Schmidt participating. General Intuition trains AI to understand and act in the physical world by learning from video-game recordings.

Most videos teach a model what the world looks like. To interact with the physical world, robots also need to learn how their actions change it. Gameplay clips also record what the player did, the exact button pressed, and the instant it happened. That pairing of image and action is the dataset an agent or a robot needs to learn cause and effect, and it barely exists outside games.

The company was spun out of Medal, a clip-sharing app with 17 million monthly users and hundreds of millions of hours of recorded gameplay. That library served as the initial dataset for training General Intuition’s model.

General Intuition builds what it calls action models and world models: systems that learn how actions change the world and can generate new environments for practice. They demonstrated that the same model playing 100 hours of Fortnite was also the one controlling a robot moving around the office.

General Intuition co-founder Pim de Witte says that “text compresses the four dimensions of reality into a single dimension”. Embodied interactions based on video games are what they believe is missing. Most of the $320 million will go toward rented compute to train the next version.

2) OpenAI Unveils Its First Custom Inference Chip

On June 24, OpenAI and Broadcom unveiled Jalapeño, OpenAI’s first custom chip. It is built for AI inference, the process of running a model after it has already been trained. It is an ASIC (Application-Specific Integrated Circuit), meaning it is hard-wired for one job rather than a general-purpose GPU.

It went from first design to production in nine months, the fastest cycle OpenAI and Broadcom know of for a chip this advanced. OpenAI used its own models to help design the chip: better models help design better chips, and better chips help run better models. Initial deployment is scheduled for the end of 2026.

A chip tuned to OpenAI’s own inference can beat a general-purpose GPU on power and price. Early testing shows performance per watt well above current state-of-the-art parts. Broadcom and Celestica build the silicon while OpenAI designs it around LLM fundamentals. This shows OpenAI clearly moving down the AI stack, expanding its moat by owning more of the infrastructure behind its core product.

A day later, IBM announced the world’s first sub-1-nanometer transistor chip. The scale would be 10,000 times smaller than a human red blood cell. IBM’s new chip is able to pack 100 billion transistors onto a chip the size of a fingernail. The breakthrough was supported by a new transistor architecture, called nanostack, a three-dimensional, nanosheet-based design. IBM also demonstrated a 40 percent scaling improvement in SRAM, the ultra-fast memory inside a processor. Because SRAM helps move data quickly inside the chip, expanding it could reduce bottlenecks in heavy computing tasks like AI. IBM sees a path to production within the next five years.

3) Meta Spends $900M to Fix WhatsApp Pay

CRED is an Indian fintech that began as an app for paying credit-card bills and now offers payments, lending, and savings to 17 million members. It handles more than 40% of India’s credit-card bill payments. On June 22, Meta invested $900 million for about 20% of the company at a $4.5 billion valuation, and hired CRED founder Kunal Shah to run WhatsApp worldwide, replacing the executive who had led it for seven years. CRED says Meta will not receive any of its customer data.

WhatsApp has more than 500 million users in India, but almost no share of the money they move. UPI, India’s instant bank-to-bank payment network, ran 23.2 billion transactions in May, and WhatsApp Pay only handled 0.65% of them. Meta has spent six years and more than $6.6 billion trying to change that. A national rule is meant to cap any single app at 30% of UPI volume. Its deadline keeps getting pushed, now to December 2026. That has allowed the current industry leaders, PhonePe and Google Pay, to remain at 46% and 33%. If the cap is ever enforced, both have to give up market share, creating an opening for WhatsApp Pay.

Learn With My Friends and Me…

Socialists Sweep NYC, China Catches Up in Coding, AI Memory Crunch, Micron’s Blowout Quarter

Other Reading…

• OpenAI Leans Toward Waiting Until Next Year for I.P.O. (NYT)

• Words vs. Worlds (Science)

On X…

Each is betting that the best way to reverse aging is with cellular reprogramming: the process of resetting cells without changing their identity.

The theory is that aging is partly driven by changes in how cells read and use their DNA, so resetting those instructions could restore function across many tissues at once rather than treating one age-related disease at a time.

But there are significant risks with this approach.

Most cells in the human body are specialized. They have specific jobs: skin cells, muscle cells, and neurons, for example. Fully reprogramming a cell means removing that cell’s specialization. When this happens, the cell abandons its core function and becomes stem cell-like.

But if every cell in a tissue is reset to a stem cell-like state, the tissue can no longer perform its normal function. Full reprogramming can also trigger uncontrolled growth and tumor formation.

Yet within the risks of full reprogramming was a breakthrough that earned stem-cell researcher Shinya Yamanaka the Nobel Prize in 2012.

Yamanaka and his team showed that four proteins, now called Yamanaka factors, when introduced into an adult cell, act as a master reset: they switch off the cell’s specialized identity and return it to the blank, embryonic-like state it started from.

It was later demonstrated that if cells are exposed to these Yamanaka factors for only a short amount of time, they can improve cellular function and tissue repair without erasing their specialization. In other words, an old cell can regain some functions lost with age, with a lower risk of tumor formation.

This is called partial reprogramming.

Animal studies suggest partial reprogramming can speed up muscle regeneration, reduce scar formation, and reduce vision loss once thought permanent.

Partial reprogramming is just one approach within Advanced Bioengineering, itself one of four scientific paths seeking to cure aging (more in our deep dive below).

Every lab, every startup, every person writing a check in longevity biotech, is making one of these four bets. You can sort them with two questions:

1. Does the approach try to buy time, or stop and reverse aging outright?

2. Does the approach require us to understand what causes aging, or can it work without that knowledge?

Each strategy has different requirements for success, regulatory paths, capital structures, timelines, and risks.

To understand these bets fully, Social Capital partnered with a team of leading researchers who are actively developing and testing these approaches to produce a two-part deep dive into longevity biotech and aging research.

In part one, we focused on the biology of aging, why it is the single biggest risk factor of death, and the structural bottlenecks holding back innovation in the space.

In part two, we focused on the four approaches researchers and companies are using to overcome these bottlenecks and answer the following questions:

1. What does each strategy try to achieve, and what is the evidence behind them?

2. Which of these four bets is closest to producing an approved therapy, and which is furthest?

3. What are the first clinical milestones each strategy can reach?

4. Where is capital flowing?

Special thanks to Aleksandr Sviridov and his colleagues, whose research contributions helped shape this Deep Dive.

Sign up below to read the full Deep Dive (and get all our past releases, including part one of this series).

After you read, let me know which strategy you think has the most promise and where you’d like us to go deeper next.

Chamath

Disclaimer: The views and opinions expressed above are current as of the date of this document and are subject to change without notice. Materials referenced above will be provided for educational purposes only. None of the above will include investment advice, a recommendation or an offer to sell, or a solicitation of an offer to buy, any securities or investment products.

Deep Dive PDF below ↓

Read more

Caught My Eye…

1) Chinese Open Model Reaches the AI Frontier at One-Sixth the Price

On June 16, the Beijing lab Z.ai released GLM-5.2, a frontier model with open weights anyone can download under an MIT license, and a million-token context window.

It beats GPT-5.5 on several long-horizon coding benchmarks at roughly a sixth of the price. This means on some benchmarks, the best model available to the public is an open one from China.

Its efficiency comes from its architecture and training. GLM-5.2 is a mixture-of-experts model: about 744 billion parameters in total, but only roughly 40 billion fire for any given token, so it answers at the cost of a far smaller model. It also uses a sparse-attention scheme borrowed from DeepSeek that lets it read a million tokens of context without the bill scaling with every word.

Additionally, they organized the post-training run around agentic engineering, getting the model to write and debug its own code over long stretches rather than only answer questions.

On the Artificial Analysis Intelligence Index, it sits #3, making it the best open weight model on the market today. Its stock on the Hong Kong exchange continues to hit new all-time highs since its IPO in Jan 2026.

Elon believes China will be able to develop a Mythos-class model by Q1 2027, and Tang Jie, the founder of Z.ai, says it will come sooner.

Enterprise buyers are already reaching for cheaper open models as token costs ramp up while quality gaps close. On June 16, Axios reported that Microsoft is weighing a self-hosted, fine-tuned version of DeepSeek V4, another Chinese open model, as a lower-cost engine for Copilot Cowork.

2) Midjourney Is Building a 60-Second Alternative to MRI

On June 17, Midjourney, the AI lab known for generating images from text prompts, announced a medical division and a machine that scans the whole body in minutes.

You stand on a platform that then lowers and submerges into water. It is lined with roughly 500,000 sound emitters, each the size of a grain of sand, that map soft tissues and bones. No radiation or magnets. The output is a 3D map of soft tissue and bone down to half a millimeter in resolution, with MRI-like quality.

A full-body MRI takes 60 to 90 minutes; Midjourney’s first prototype currently takes about 20 minutes, with a goal of reaching 60 seconds.

The ring is lined with roughly 500,000 ultrasound sensors built on Butterfly Network’s computer chips. Each sensor sends out sound waves and records the echoes that reflect off the body, generating terabytes of raw data. The challenge is reconstructing those echoes into a detailed 3D image. Processing the roughly 17 gigabytes of data generated every second requires 21 servers providing about two petaflops of computing power.

Midjourney is in a unique position with no outside investors and was profitable within months of its public launch in 2022. It surpassed $500m in ARR in 2025.

The goal is to build 50,000 scanners doing a billion scans a month by 2031, installed in spas where the scan is a side effect of a soak. The first opens in San Francisco next year. The FDA path also starts narrow, with body composition maps first rather than diagnostic medical capability. They will submit regular test results to the FDA to get approval for more advanced capabilities as they advance their hardware.

3) Anduril Wins the Fastest Fighter Production Contract in 50 Years

On June 17, the Air Force moved its Collaborative Combat Aircraft program into production, awarding contracts to Anduril and General Atomics to build their uncrewed fighters. Collaborative Combat Aircraft are semi-autonomous jets designed to fly alongside crewed fighters like the F-35, carrying weapons and sensors at a fraction of the cost of a piloted plane.

Anduril’s aircraft FQ-44 drops its prototype label and enters serial production. The production line at Arsenal-1 in Ohio can already produce up to 150 aircraft a year, with modular tooling that allows for expansion and iteration.

Anduril is nine years old, and this is the first time a new company has won a U.S. fighter program since the 1970s. It went from prototype award in April 2024 to first flight last October to a production contract this month, representing the fastest path for a fighter in more than 50 years.

The contract lands a month after Anduril raised $5 billion at a $61 billion valuation, double its worth a year earlier.

4) Fox Buys Roku for $22 Billion

On June 15, Fox agreed to buy Roku for about $22 billion, paying $160 per share, a 34% premium over Roku’s June 11 close.

Roku is the connected-TV platform that sits on more than 100 million households’ screens and runs The Roku Channel. Fox already owns Tubi, which means it now has two industry leaders of the free ad-supported streaming television (FAST) market. The deal is set to close in the first half of 2027, pending antitrust review and a shareholder vote.

What Fox is buying is the home screen and the data behind Roku. Roku is the most-used TV operating system in the US, ahead of Samsung. Roku can also see everything that plays on the screen, every show, ad, and app, from any source, through automatic content recognition. That viewing data is what makes its ad targeting valuable.

The purchase reverses a trade Fox made six years ago, when it sold its 5% stake in Roku to help pay for Tubi. Now it owns the whole platform. The combined company would rank third in US television by share of viewing, just behind Disney and YouTube.

Learn With My Friends and Me…

All-In: World’s First Trillionaire, Anthropic Fable Banned, The New Oligarchs, Iran Peace Deal

Other Reading…

• Agentic Coding and Persistent Returns to Expertise (Anthropic)

• Dear A.I. Companies: The Doom Trolling Needs to Stop (The New York Times)

• Block rolls out Builderbot (Block)

On X…

In his first decision as Chair, Warsh held the benchmark rate at 3.5–3.75% in a unanimous vote, but the Fed’s own projections now show nearly half the committee expecting a hike before year-end, as seen in the chart below.

Warsh abstained from submitting a dot, signaling a potential change to the future Summary of Economic Projections (SEP).

Understanding why any of this matters starts with understanding how the Federal Reserve actually works.

You probably have a vague sense of what the Fed does.

They raise and lower rates. They focus on employment. But most people, even financially literate ones, could not tell you why it exists, how it actually works, or what it is really trying to do. So let’s examine it from the ground up.

The Federal Reserve controls the cost of borrowing in the United States.

It does this by setting the federal funds rate: the rate banks charge each other to borrow money overnight. That rate anchors the cost of money across the financial system.

Treasury yields and other market rates move in response to the current rate and where investors expect it to go. Lenders then add a premium for time, credit risk, and profit, producing the rates charged on mortgages, corporate loans, credit cards, and nearly every other form of borrowing.

When the federal funds rate rises, borrowing becomes more expensive. Households spend less, companies invest and hire less, and investors place a lower value on future earnings. Bond prices, stock valuations, real estate values, retirement accounts, and the strength of the dollar all adjust in response.

This makes the Fed the most consequential institution in the American economy.

Kevin Warsh is now in charge of it, and he has a clear vision for how it should change.

In April Senate Banking Committee testimony, Warsh articulated a leaner central bank: one that acts less, telegraphs less, and treats market surprise as a feature rather than a flaw. No Fed Chair in the last two decades has operated this way.

If that vision were to become a reality, it would reshape how the Fed interacts with markets across four main dimensions:

As the Fed communicates less and surprises more, investors have less certainty about what comes next. Less certainty means higher risk premiums, which could push long-term bond rates up and stock valuations down, even without a single rate hike.

That is the direction Warsh wants to take the Fed. Here is his starting point:

• CPI inflation at 4.2% year-over-year, a three-year high, driven by an energy-price spike, well above the Fed’s 2% target

• Unemployment at 4.3%, steady in the 4-5% range over the past two years

• Jerome Powell, the outgoing Fed Chair, is staying on the board until 2028, retaining a vote on every rate decision Warsh makes

To see where the Fed goes next, we need to layer in the forces shaping its rate decisions: its history, its private-public hybrid architecture, its modern toolkit, and the political and market pressures surrounding every decision.

My research team traced the Fed from its creation to the institution Kevin Warsh now inherits. The result is a 108-page deep dive on how it was built, how power moves through it, and whether it can remain independent under today’s political and market pressure.

Inside, we’ll cover:

1. Why did America spend its first century debating whether a central bank was unconstitutional, and what finally broke that resistance in 1913?

2. If the President can’t fire the Fed Chair, why is the Federal Reserve vulnerable to political pressure?

3. Every Fed Chair has faced a defining test. What were those tests, how did Warsh’s predecessors respond, and what will define his tenure?

4. How do markets respond after every Fed Chair transition in history?

Read the deep dive below and let me know what you think.

Chamath

Disclaimer: The views and opinions expressed above are current as of the date of this document and are subject to change without notice. Materials referenced above will be provided for educational purposes only. None of the above will include investment advice, a recommendation or an offer to sell, or a solicitation of an offer to buy, any securities or investment products.

Deep Dive PDF below ↓

Read more

Caught My Eye…

1) OpenAI third phase: an AGI for everyone

On June 8, OpenAI submitted a confidential draft S-1 to the SEC, announcing the filing itself. The filing came one week after Anthropic disclosed its own IPO preparations. OpenAI was last valued at $852 billion in March; Anthropic closed its Series H at $965 billion on May 28.

The same day, Sam Altman and Jakub Pachocki published “Built to benefit everyone,” declaring a “third phase” for the company built around three goals: (1) an automated AI researcher, with ‘a significant fraction’ of OpenAI’s research done by AI by March 2028; (2) faster economic growth; and (3) a personal AGI for everyone on Earth. Their first phase was a research company; their second was a product company. Now they want to expand into a broader distribution company that ensures everyone can thrive with these tools.

The WSJ reported OpenAI is weighing sharp cuts to token prices in anticipation of similar cuts at Anthropic. The same week, SemiAnalysis measured how much usage the flat-rate AI plans allow: a $ 200-per-month ChatGPT Pro subscription can draw up to $14,000 in tokens per month at API prices; Claude Max, $8,000. SemiAnalysis pegs the cost of serving tokens at roughly a quarter of API list price, in line with the ~75% inference margins both labs are reported to earn. At that rate, a maxed-out Pro user costs OpenAI about $3,500 in compute against $200 in revenue.

And on June 11, OpenAI agreed to acquire Ona, whose customer-controlled cloud environments let Codex agents keep working for hours or days inside a company’s own infrastructure. Codex now counts 5 million weekly users, up 400% this year.

2) Bezos takes a CEO seat to build an artificial general engineer

On June 11, Project Prometheus emerged from stealth with a $12 billion Series B at a $41 billion valuation. The company is seven months old, employs roughly 150 people across San Francisco, London, and Zurich, and is co-run by Jeff Bezos and Vik Bajaj, the Google X veteran who co-founded Alphabet’s Verily, an artificial intelligence and precision health company. It is Bezos’s first CEO seat since he left Amazon in July 2021.

Prometheus is building what Bezos calls an “artificial general engineer”: AI tools that compress the design-to-manufacturing cycle for physical objects, from bridges to chips, which he described as a very modern version of CAD that makes the build loop 10 times faster. The constraint it claims to solve is data. Language models were trained on the internet; engineering AI needs physics-grounded simulation and proprietary test results that largely don’t exist online. Prometheus generates most of that data itself, supplemented by results from manufacturers. Target industries are computing, aerospace, automotive, advanced manufacturing, and drug discovery.

Bezos believes AI will bring “labor scarcity” rather than mass unemployment, his term for a world where productivity lets companies take on more projects than they have people to staff. He also believes that with a higher standard of living and productivity, people can choose to work less or scale down their workload.

3) SpaceX reveals AI1 data center satellite

On June 8, three days before pricing its IPO, SpaceX revealed AI1, its first orbital data center satellite. The spacecraft targets 120 kW of sustained compute (150 kW peak) at roughly 600 km altitude, with a 70-meter wingspan, wider than a Boeing 747-8, and a 110-square-meter liquid radiator to dump waste heat into space. The compute payload is interchangeable, so chips can be swapped as hardware generations turn over. Two prototypes are slated for early 2027, and SpaceX has filed plans for a constellation of up to 1 million satellites.

On Friday, June 12, SpaceX began trading on Nasdaq as SPCX in the largest IPO in history. Priced at $135 a share the night before, a $1.77 trillion valuation that raised about $75 billion, the stock opened at $150, ran as high as $176.52, and closed up roughly 19% near $161, pushing SpaceX past a $2 trillion valuation on its first day and making it the sixth-largest U.S. public company.

The heavily oversubscribed deal more than doubled Saudi Aramco’s $29.4 billion fundraising record. SpaceX says proceeds will fund AI compute infrastructure first, then launch vehicles and satellite constellations.

To get a full picture of SpaceX, we just published our 120-page SpaceX Deep Dive, covering everything from launch economics to the orbital compute thesis and more.

4) Ramp’s $44B bet on metered intelligence as corporate spend

On June 4, Ramp raised $750 million at a $44 billion valuation, nearly triple its value a year ago, on more than $1 billion in annualized revenue with positive free cash flow. Ramp is a corporate spend platform (cards, expenses, bill pay, and procurement) serving more than 70,000 businesses. The core thesis is that AI tokens have become a major corporate cost that current financial systems can’t model well. Ramp is now betting that helping companies measure and control those costs will open up a new revenue opportunity.

CEO Eric Glyman’s accompanying essay, “The Third Pillar,” argues that business spending ran on two pillars, people and vendors. Metered intelligence is now the third.

Their new early-access product pulls token-level usage data from Anthropic, OpenAI, Gemini, and Cursor into a single dashboard, attributes it to teams and use cases, and recommends routing work to cheaper models. The spread it monetizes is wide: GPT-4-level intelligence has fallen from $60 to roughly $0.40 per million tokens since 2023, while businesses rely on more expensive frontier models for menial tasks. Rerouting 10% of a $10 million AI bill from frontier to commodity intelligence saves nearly $1 million.

Learn With My Friends and Me…

Anthropic's Fable Backlash, Nationalizing AI, Inflation Heats Up & California’s Broken Elections

Other Reading…

• China’s Unitree Will Dominate Global Robotics (SemiAnalysis)

• A Manufacturing Revolution Is Happening—in Space (The Free Press)

• Policy on the AI Exponential (Dario Amodei)

On X…

That valuation is based on the thesis that SpaceX has opened the lowest-cost route into space, and that the largest opportunities are still ahead.

What is that opportunity, and can SpaceX capture it?

One useful analogy starts in 1497.

For centuries, spices reached Europe over land, via routes controlled by Arab and Venetian middlemen. By the time pepper arrived, the price had been marked up 20-30 times.

Vasco da Gama, a Portuguese mariner and explorer, found a sea path around the Cape of Good Hope, bypassing every middleman on that land route entirely.

And that single move collapsed the entire cost structure of the trade. The companies that controlled the new maritime trade became some of the most valuable in history.

England chartered the East India Company in 1600 to sail the route itself. The Dutch followed with the Dutch East India Company (VOC) in 1602.

At its peak in 1720, the VOC’s market capitalization reached roughly a third of the Dutch Republic’s entire annual economic output, a share that dwarfs every company in existence today, including Nvidia and Apple.

SpaceX is changing the economics of reaching space the way Vasco da Gama changed the economics of reaching Asia.

For European merchants, the prize was spices. For the modern economy, space opens three potential prizes:

1. Connectivity. Most of the world’s internet still depends on physical cables, and those cables do not reach everywhere. Starlink can beam broadband from orbit to anyone, anywhere.

2. Compute. AI data centers on Earth are hitting a power wall. A new data center needs a grid connection that takes years to permit, and electricity is getting scarcer and more expensive. Space offers the potential of abundant solar power without the same land, grid, and permitting constraints that data centers face on Earth.

3. Critical minerals. The rare earth elements that AI chips and energy storage depend on are largely controlled by China. The Moon and the asteroid belt hold reserves that dwarf anything on Earth, outside any single country’s supply chain.

All of it sits behind one barrier: the cost of reaching orbit.

For sixty years, reaching orbit cost so much that only governments could afford it. But SpaceX changed that.

In 2025, they launched 85% of all satellites, more than every other space program on Earth combined. They did it by making rockets reusable, increasing launch cadence, and bringing more manufacturing in-house.

NASA’s Space Shuttle cost $54,500 per kilogram to reach low Earth orbit. SpaceX drove that number down to $2,720, a 20x reduction, documented in NASA’s own cost analysis. (The deep dive below breaks down exactly how they did it.)

SpaceX now sits on the main commercial route to orbit, and Starlink is the first proof of what that route can unlock.

Cheap launch let SpaceX build the world’s largest satellite internet network and refresh it faster than anyone else. Starlink turned access to orbit into $11.4B of recurring broadband revenue in 2025.

The same route that made orbital broadband possible may now reveal the next frontier opportunity for SpaceX: orbital compute.

That is the opportunity we wanted to understand. My research team at Social Capital spent months speaking with industry leaders and investors, reviewing hundreds of hours of material, and stress-testing the SpaceX story from the ground up.

We put it all into a 120+ page Deep Dive: what SpaceX has built, what this route can unlock next, and where the constraints remain.

A few things we cover inside:

1. How SpaceX went from near-bankruptcy in 2008 to outlaunching every space program on Earth, and why that lead is nearly impossible to replicate.

2. What SpaceX makes possible once launch becomes cheap, reusable, and routine, and how satellites can compete for the $2.1T telecom market.

3. Starlink: the satellite business posting software-grade margins inside an aerospace company

4. Orbital compute: the xAI merger, the AI-1 satellite, and the next bet the valuation is pricing in

Read the deep dive and let me know what you think.

Chamath

Disclaimer: The views and opinions expressed above are current as of the date of this document and are subject to change without notice. Materials referenced above will be provided for educational purposes only. None of the above will include investment advice, a recommendation or an offer to sell, or a solicitation of an offer to buy, any securities or investment products.

Deep Dive PDF below ↓

Read more

Caught My Eye…

1) Counterfeit Bug in Zcash found via Opus 4.8

On May 29, security researcher Taylor Hornby disclosed a critical bug in Zcash’s Orchard shielded pool, the privacy coin’s most advanced transaction layer. The flaw, present since Orchard launched on May 31, 2022, would allow an attacker to generate counterfeit ZEC within the pool without leaving any visible on-chain signature. Hornby found it using Anthropic’s Opus 4.8 model, running inside a custom audit-agent framework he built. What took years of human audits to miss, a new AI model found within 28 hours of release.

Orchard’s privacy works through a zero-knowledge proof: the spender mathematically proves they own a coin and are spending it for the first time, without revealing which coin. To prevent double-spending, every spend publishes a one-time tag called a nullifier, derived from the owner’s secret key. Each coin should have exactly one valid nullifier. The bug skipped one step of math that ties the secret key to the coin’s public address, allowing a malicious user to generate a different nullifier for the same coin each time they spend it. Hornby’s proof of concept doubled a single note’s balance repeatedly until his test wallet held 10 million ZEC, and every proof remained valid.

The Zcash Open Development Lab shipped an emergency soft fork on June 2, disabling Orchard, then activated a network upgrade on June 3, re-enabling the pool with the corrected circuit. Currently, there is no evidence of exploitation, but because Orchard hides every transaction by design, the bug is difficult to audit after the fact. No one can reconstruct the shielded history to confirm it was never used. On June 6, Shielded Labs, the Zcash Foundation, and ZODL proposed Ironwood, an upgrade that would launch a new shielded pool. Anyone running a node could then sum balances across pools and verify the supply is clean.

The Orchard flaw is the second critical Zcash bug disclosed in 90 days. The first came in March, when researcher “Sol” found a separate flaw in the legacy Sprout pool, where node software was skipping zero-knowledge proof verification under certain conditions, which had been dormant since July 2020.

2) Fusion’s First Commercial Construction Site

On June 4, Helion Energy raised $465 million in Series G funding at a $15.5 billion post-money valuation. Total raised to date: $1.5 billion. The round funds the construction of Orion, Helion’s first commercial fusion power plant, now being built in Washington. The plant is contracted to deliver electricity to Microsoft starting in 2028.

Helion is the Sam Altman-backed fusion company pursuing a magneto-inertial approach: two field-reversed plasma rings are accelerated to over 1 million mph from opposite ends of the machine and collide in the middle, briefly reaching the temperature needed to fuse deuterium and helium-3.

In February, Polaris became the first privately funded machine to operate on deuterium-tritium fuel, reaching a plasma temperature of 150 million degrees Celsius, roughly 10 times hotter than the core of the Sun.

Helion is the first to break ground on a commercial site with a binding power-purchase agreement. They are still working on sustained net energy gain and continuous grid output with a 2028 timeline. This is the most aggressive date in the industry by roughly five years.

3) Berkshire Pays $6.8 Billion for Taylor Morrison

On June 1, Taylor Morrison agreed to be acquired by Berkshire Hathaway at $72.50 per share in cash, a 24% premium to the May 29 close of $58.50. Total equity value is $6.8 billion, and enterprise value is $8.5 billion. This is Greg Abel’s first multibillion-dollar acquisition since taking over from Warren Buffett.

Taylor Morrison is one of the top five US public homebuilders, operating across the Sun Belt and the Mountain West. Berkshire already owns Clayton Homes in manufactured housing and Berkshire Hathaway HomeServices in brokerage. Adding Taylor Morrison gives Abel a site-built builder.

“Over time, we expect to unify our site-built homebuilding operations into a combined platform enabling us to deliver the dream of homeownership to more Americans.” - Greg Abel, President and CEO of Berkshire Hathaway.

Buffett told CNBC that “Greg did that faster than I could have done it, smoother than I could have done it, and I never talked to the CEO… He has launched.” Berkshire continues to hold a record $397 billion cash pile.

Learn With My Friends and Me…

Other Reading…

• How AI Makes Initiative Beat Intelligence (Tyler Cowen)

• When AI Builds Itself (Anthropic)

• Will You Beat Google? (Sreeram Kannan)

On X…

It is the single largest risk factor for many of the diseases that kill people in the modern world, ahead of smoking, blood pressure, or cholesterol. Cure cancer outright, and average life expectancy rises only about three years, because the underlying risk keeps compounding.

This is one of the premises behind longevity biotech: if aging is the one biological process that drives many diseases at once, then slowing it could delay all of them together, and let people live longer in good health.

The economic incentives are there as well. Slowing human aging by a single year has been estimated to add about $400B a year to U.S. GDP.

Animal evidence shows aging can be slowed. Calorie restriction has consistently extended lifespan in mice, with evidence dating back to 1935. In one mouse study, animals on the strongest calorie-restricted diet lived roughly 50 to 65% longer than controls.

A single gene change, or a drug like rapamycin started late in life, can do the same. And a few species, from the naked mole rat to some turtles, barely age at all, their risk of death holding flat as the years pass.

In humans, the evidence is thinner. A few common drugs, statins and the new GLP-1 medicines among them, already lower overall mortality in specific patient groups, but cutting deaths from one disease is not the same as slowing aging, and nothing has been approved to do the latter. Proving that one slows aging would mean a trial running for decades, outlasting its own patent, with no one able to profit from the result. And no regulator yet treats aging as a condition for which you can be prescribed something.

The money follows that logic. Research into the biology of aging receives $1-2B a year worldwide, ~0.25% of $350–400B in global biomedical R&D, even though it sits upstream of the diseases that consume most of the developed world’s health spending.

Aging is the largest shared risk factor of death, and the one we spend the least to understand. The therapy that could matter most is the hardest to ever prove. Whether aging is a problem we can engineer our way out of, or a biological limit we can only manage, is now one of the most consequential open questions in medicine.

Our Social Capital team and industry research partners examined longevity biotech and aging research from first principles, and the 100+ page Deep Dive below maps our findings, including:

• The two opposing theories of what aging is, and what each implies for treating it

• A century of animal evidence on lifespan extension

• What affects aging in humans

• Why the 2014 metformin longevity finding does not hold up

• The five interlocking causes of the funding gap afflicting aging research

Special thanks to Aleksandr Sviridov (@lex_sviridov) and his colleagues, whose rigorous research contributions helped shape this Deep Dive.

If you want to learn about the field of longevity with me, sign up below to read the full Deep Dive (and get all our past releases).

Chamath

Disclaimer: The views and opinions expressed above are current as of the date of this document and are subject to change without notice. Materials referenced above will be provided for educational purposes only. None of the above will include investment advice, a recommendation or an offer to sell, or a solicitation of an offer to buy, any securities or investment products.

Deep Dive PDF below ↓

Read more

Caught My Eye…

1) Anthropic Raises $65B at $965B, Passing OpenAI

On May 28, Anthropic announced a $65 billion Series H at a $965 billion valuation. The round more than doubles the company’s $380 billion mark from February and clears OpenAI’s $852 billion close in late March. It’s the first time Anthropic has been valued above its primary frontier competitor, OpenAI.

The same day, Claude Opus 4.8 shipped at the same $5 / $25 per million token price as 4.7. SWE-bench Pro climbed to 69.2% from 64.3%, and Anthropic claims the model is roughly four times less likely to leave bugs in its own code unremarked. Their evaluations also showed that Opus 4.8 is the first Claude model to get a perfect score on their benchmark that tests for lazy thinking and work. This means instead of cutting corners and making assumptions on unintuitive and poorly documented code bases, it will trace things down fully to avoid significant misunderstandings.

Anthropic says Mythos-class capability will reach all customers “in the coming weeks,” and added that no lab has yet built safeguards strong enough to prevent misuse.

2) Pluralis Trains a 7.5B Model Across 198 Cities

Pluralis Research, a research lab focused on collectively-owned AI, previously completed Node0-7.5B, a 36-billion-token pretraining run executed over three weeks using 1,642 GPUs across 198 cities. Most contributions came from consumer-grade Nvidia 4090s and 6000-series cards. Founded in 2024 by ex-Amazon ML scientist Alexander Long, Pluralis raised a $7.6M seed led by USV and CoinFund, with Variant, Topology, Bodhi Ventures, Balaji Srinivasan, and Clem Delangue participating.

Pluralis recently went live with Pluralis-8B on their new Agora system, an 8-billion-parameter language model being trained across roughly 50 cities on a mix of consumer GPUs and small datacenters connected over ordinary internet.

The company’s approach is called Protocol Learning. They split the model into pieces distributed across participants such that no single party can reassemble the full weights, while the math still produces a working model whose inference revenue flows back to contributors. The mechanism, described in NeurIPS 2025’s “Unextractable Protocol Models“ paper, periodically scrambles the seams between those pieces in a way that preserves model behavior but blocks reconstruction.

EpochAI has calculated that the growth rate of frontier AI training is 5x/year. In comparison, they measure that decentralized runs are scaling at 20x per year. If both trends hold, it would take five and a half years for decentralized training to catch up.

“Pluralis paves the way to true collective ownership at the model layer, by training foundation models that are split across geographically separated devices connected only over the internet. It guarantees public access to strong base models, enables anyone to participate in the economic benefits of AI, and unlocks open innovation. There may be no technical challenge today that has a larger impact.” - Alexander Long, Founder and CEO, Pluralis Research

3) Quantinuum Sets $1B IPO Terms at $12.7B Valuation

On May 26, Quantinuum filed an amended S-1 setting the price range for its Nasdaq IPO at $45 to $50 per share on 21 million Class A shares. The deal targets up to $1.05 billion in proceeds at a $12.7 billion fully diluted valuation, with trading expected on or around June 3 under ticker QNT. It is the first traditional IPO from a full-stack quantum computing company, while every prior public-market quantum entry has come through a SPAC.

Quantinuum builds quantum computers using individual trapped ions as qubits, a different physical approach from the superconducting-circuit machines that Google and IBM run. The company was formed in November 2021 when Honeywell spun out its internal quantum hardware division and merged it with Cambridge Quantum, a UK quantum-software firm. Honeywell has remained the majority shareholder and has led multiple follow-on funding rounds. Its current commercial system, Helios, runs 98 physical qubits and 48 logical qubits, the error-corrected ones that actually run useful programs. Their launch customers include Amgen, BMW, JPMorgan Chase, and SoftBank. After the IPO, Honeywell will retain about 49.1% of voting power.

2025 revenue came in at $30.9 million, up from $23 million the year prior, against a $192.6 million net loss. The company has spent more than $2 billion on R&D over the past decade and held $677 million in cash at quarter-end.

Learn With My Friends and Me…

Anthropic’s Digital God, Pope vs AI, Job Loss Narrative Flips, Open Source Crackdown Coming?

Other Reading…

• Avoiding Death on the Yellow Brick Road (Joe Schmidt)

• Rogo’s Big Finance Bench (Rogo)

• Pluralis: The Last Revolutionary AI Protocol (kelxyz_)

On X…