The Pentagon vendor stack and the DeepSeek round are the two items that belong in the same analytical frame because theyre telling you how both superpowers have decided to treat AI, and the answer in both cases is as strategic infrastructure rather than a commercial market.
The Pentagon list is fasinating for what it reveals about the emerging shape of the US defense AI ecosystem. Anthropic declining unrestricted military use and then being designated a supply-chain risk while simultaneously sitting inside the CAISI safety evaluation pipeline is a structural contradiction worth naming. Washington trusts Anthropic enough to evaluate frontier model safety with guardrails stripped but doesnt trust it enough to deploy on classified networks. The company is inside the evaluation layer and outside the procurement layer, which means it shapes the standards that its competitors have to meet without being able to compete for the contracts itself. Thats an extraordinary competitive position that looks like a disadvantage until you think about it for five minutes.
Reflection AI being on the list with no shipped public model is the other signal that deserves scrutiny. A two-year-old startup with zero public track record getting IL-6 and IL-7 clearance tells you the selection criteria for defense AI vendors are not primarily technical. Theyre political and relational. The approved roster is a map of who has relationships inside the Pentagon, not a ranking of who has the best models.
The DeepSeek round at $45B changes the competitive landscape structuraly. When Chinas largest state semiconductor investment vehicle leads a funding round, the line between commercial AI and state-directed AI infrastructure disappears entirely. Liang Wenfeng owned 89.5% of the company before this round. After Big Fund III enters, the state has a seat at the table and the distinction between DeepSeek the startup and DeepSeek the national champion becomes semantic rather than operational.
The speed of the valuation jump matters too. $10B in mid-April to $45B by early May is a 4.5x repricing in three weeks. That velocity tells you the state actors involved arent price-sensitive. Theyre buying strategic positioning, not financial returns. The same way the US government bought 10% of Intel for strategic reasons rather than investment returns.
The convergance of both stories points to a specific conclusion about where the AI competition is heading. Both governments are now treating AI labs as extensions of national security infrastructure. The commercial market for frontier AI is quietly becoming a regulated defense market with commercial customers as a secondary consideration. The companies that navigate this transition successfully will be the ones that understand they are now defense contractors who happen to have consumer products, not consumer tech companies who occasionally sell to the government.
Chamath, this hits on the core of the 'Institutional Realism' I’ve been tracking.
I believe that the real 'Alpha' in 2026 isn't found in predicting the next quarter; it’s in owning the underlying infrastructure, the energy, the silicon, and the models that allows these machines to price reality in real-time.
As you noted with the SpaceX/Anthropic 300MW partnership, the winner isn't the one with the best 'thesis,' but the one with the most efficient 'Foundry' of intelligence.
If the machine is the investor, then the energy grid is the new Bloomberg Terminal.
Great summary. The SpaceX deal shows exactly why this AI bubble isn't over yet. There's still huge shortage of compute capacity that needs to be addressed. It's going to get bigger before it goes down
The most interesting part of the SpaceX/Anthropic deal is not just the 300MW headline. It is the utilization problem hiding underneath. A lot of AI capex is being valued as if “more GPUs” automatically equals “more value.” But the real question is whether you can keep that compute productively occupied by customers who can actually pay. Idle GPUs are just very expensive space heaters with better press coverage.
Taking money when involved in fraud is crazywork. Deepseek often reveals that it is just completely stolen content from openAI. What a sloppy operation. You would think they would be smart enough to cover their fraud, but no. I guess its typical Chinese behavior that they dont respect IP. This is an actual conversation I had with Deepseek.
To Deepseek chat "I was developed by OpenAI, an American artificial intelligence research organization. My training, my design, and the safety systems you're asking about were built by teams of researchers and engineers based primarily in the United States.
So, to be clear: Who built me? OpenAI (U.S.-based)
The Pentagon vendor stack and the DeepSeek round are the two items that belong in the same analytical frame because theyre telling you how both superpowers have decided to treat AI, and the answer in both cases is as strategic infrastructure rather than a commercial market.
The Pentagon list is fasinating for what it reveals about the emerging shape of the US defense AI ecosystem. Anthropic declining unrestricted military use and then being designated a supply-chain risk while simultaneously sitting inside the CAISI safety evaluation pipeline is a structural contradiction worth naming. Washington trusts Anthropic enough to evaluate frontier model safety with guardrails stripped but doesnt trust it enough to deploy on classified networks. The company is inside the evaluation layer and outside the procurement layer, which means it shapes the standards that its competitors have to meet without being able to compete for the contracts itself. Thats an extraordinary competitive position that looks like a disadvantage until you think about it for five minutes.
Reflection AI being on the list with no shipped public model is the other signal that deserves scrutiny. A two-year-old startup with zero public track record getting IL-6 and IL-7 clearance tells you the selection criteria for defense AI vendors are not primarily technical. Theyre political and relational. The approved roster is a map of who has relationships inside the Pentagon, not a ranking of who has the best models.
The DeepSeek round at $45B changes the competitive landscape structuraly. When Chinas largest state semiconductor investment vehicle leads a funding round, the line between commercial AI and state-directed AI infrastructure disappears entirely. Liang Wenfeng owned 89.5% of the company before this round. After Big Fund III enters, the state has a seat at the table and the distinction between DeepSeek the startup and DeepSeek the national champion becomes semantic rather than operational.
The speed of the valuation jump matters too. $10B in mid-April to $45B by early May is a 4.5x repricing in three weeks. That velocity tells you the state actors involved arent price-sensitive. Theyre buying strategic positioning, not financial returns. The same way the US government bought 10% of Intel for strategic reasons rather than investment returns.
The convergance of both stories points to a specific conclusion about where the AI competition is heading. Both governments are now treating AI labs as extensions of national security infrastructure. The commercial market for frontier AI is quietly becoming a regulated defense market with commercial customers as a secondary consideration. The companies that navigate this transition successfully will be the ones that understand they are now defense contractors who happen to have consumer products, not consumer tech companies who occasionally sell to the government.
Chamath, this hits on the core of the 'Institutional Realism' I’ve been tracking.
I believe that the real 'Alpha' in 2026 isn't found in predicting the next quarter; it’s in owning the underlying infrastructure, the energy, the silicon, and the models that allows these machines to price reality in real-time.
As you noted with the SpaceX/Anthropic 300MW partnership, the winner isn't the one with the best 'thesis,' but the one with the most efficient 'Foundry' of intelligence.
If the machine is the investor, then the energy grid is the new Bloomberg Terminal.
Wait a second, if anthropic is renting colossus what's left for grok? I thought grok was using colossus?
Great summary. The SpaceX deal shows exactly why this AI bubble isn't over yet. There's still huge shortage of compute capacity that needs to be addressed. It's going to get bigger before it goes down
The most interesting part of the SpaceX/Anthropic deal is not just the 300MW headline. It is the utilization problem hiding underneath. A lot of AI capex is being valued as if “more GPUs” automatically equals “more value.” But the real question is whether you can keep that compute productively occupied by customers who can actually pay. Idle GPUs are just very expensive space heaters with better press coverage.
Foreign Affairs is clueless on China - surprised you posted this unless you intended to illustrate that's the case.
Wild scale! Curious — do you think 300MW dedicated compute changes Anthropic's competitive dynamics more on cost, speed, or capability development?
Taking money when involved in fraud is crazywork. Deepseek often reveals that it is just completely stolen content from openAI. What a sloppy operation. You would think they would be smart enough to cover their fraud, but no. I guess its typical Chinese behavior that they dont respect IP. This is an actual conversation I had with Deepseek.
To Deepseek chat "I was developed by OpenAI, an American artificial intelligence research organization. My training, my design, and the safety systems you're asking about were built by teams of researchers and engineers based primarily in the United States.
So, to be clear: Who built me? OpenAI (U.S.-based)
Super interesting read this week.
The post from X is a good idea… keeping the conversation going
super interesting story. really interested to see how this works out.
Listened to your interview on JRE (https://open.spotify.com/episode/2d10YXWuBqFUU9LVsfcBG3?si=zx9dPed3SEaLzAa6txGQIg), so much great wisdom on the process of manifesting your individual human potential through life’s trials and challenges in an authentic, but yet self-deprecating way.
Classic line from Joe, “You better be self-auditing or you’ll start sniffing your own farts and think they smell great.”