The Signal
In 2022, I watched a DeFi protocol lose 40% of its liquidity providers in seven days. The treasury was audited. The roadmap was public. The token barely moved. None of that mattered. The founding team was gone, and the core infrastructure had been outsourced to a competitor. Data over drama. I reduced my position immediately and never looked back.
SemiAnalysis just published the same warning for Google DeepMind. The report’s conclusion is direct: DeepMind is no longer a frontier AI lab, and the probability of returning to state-of-the-art is zero. This is not a product review. It is an allocation report. Four senior researchers — Jeff Dean, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals — have left Google to start a new company. Gemini co-lead Noam Shazeer is now at OpenAI. Nobel laureate John Jumper is at Anthropic. In the background, the compute stack is moving in the opposite direction.
SemiAnalysis estimates that between Q3 2026 and Q4 2027, more than 20% of Google’s TPU shipments will go directly to Anthropic. That is not a cloud rental line item. That is a two-year transfer of scarce hardware to Gemini’s primary rival.
Numbers don’t lie. Liquidity vanishes. Lessons remain.
The Infrastructure Read
Let’s unpack what this actually is. DeepMind is not a random startup. It is the team behind the Transformer architecture, AlphaGo, and AlphaFold. It did not lose one founder; it lost several in one sweep. The report blames Google’s culture: bureaucratic, slow, strategically conservative. In trader terms, that means the organization has moved from growth to cash-flow management. It is prepared to earn from the market, not to dominate it.
The TPU detail is the piece most people will skip, and it is the only piece that matters. TPUs are not commodity chips. They are Google’s custom accelerators, built to train and run Google’s own models. If over 20% of TPU shipments are going to a direct competitor, the hardware is no longer an internal weapon. It is a product. In mining terms, this is the equivalent of a large BTC miner selling hashrate to a rival chain while its own chain struggles to finalize blocks.
Google still has a functional product. Gemini exists. Google Cloud is growing. But a lab that delivers updates on a schedule while the field is moving every quarter is a different animal. The report compares Google to IBM and Intel: strong technical capability, strong revenue, no appetite for the hardest race. IBM still holds patents. IBM still makes money. Nobody expects its next quarterly release to change computing history. The same is starting to be true of Google.
Order Flow Analysis
This is where order flow gets interesting. The talent exodus is not random. Four senior researchers leaving at once looks like a multi-signature unlock. That only happens when the internal debate is settled and the builders lost. The people who remain might still be brilliant, but they are not the marginal thinkers setting the next research agenda. The market’s marginal AI buyer is now at OpenAI and Anthropic.
Let me make this concrete. Gemini should benefit from Shazeer’s Mixture-of-Experts work. Now he is taking that knowledge to OpenAI. If MoE has another spring, it will be tested against Google, not for it. Jumper’s move is even more telling. AlphaFold was DeepMind’s scientific crown jewel. If the person who built it decides the next version must be built elsewhere, the research lane is not just behind schedule. It is transferring ownership.
I have seen this pattern before. Based on my audit experience in DeFi, I do not judge a protocol by token price. I judge it by resource allocation. Where is the treasury flowing? Who is getting the critical infrastructure? If a project is paying competitors to build on its stack, the architecture is already compromised. Google’s TPU sale sends the same signal at hyperscale. The parent company is willing to profit from its own challenger.
There is a nuance: what does a TPU shipment mean exactly? It means Anthropic is taking physical delivery of a chip designed and built inside Alphabet. That creates a durable counterparty relationship. Google gets revenue. Anthropic gets extremely scarce compute. But Google also loses optionality. Every TPU sitting in Anthropic’s data center is a TPU that cannot be used to scale Gemini when the next model needs it. Selling the last batch of a scarce asset is easy. Buying it back during a demand spike is not.
For years, I have used a simple three-metric system to evaluate any AI or crypto position. Talent velocity: who is leaving, who is joining. Compute retention: is the project spending its own edge on its own thesis? Execution latency: what is the gap between the roadmap and shipped code? On Google DeepMind, all three are flashing red.
So here is the core finding: Compute is the ultimate counterparty. When a lab has to rent its hardware to its direct competitor, it has already priced itself out of the race. In derivatives, you do not hedge a thesis you believe in by handing your best collateral to the person shorting you. That is not hedging. That is settlement. Google might collect interest on its position, but it no longer controls the underlying.
The Contrarian Trade
The contrarian take is that Google is not stupid. Selling TPUs to Anthropic could be the most rational capital decision available. AI has become a capex war, and models get commoditized every six months. If I were Google’s CFO, I would sell surplus hardware to anyone willing to pay. IBM and Intel did exactly this. They earn steady revenue and lose the frontier crown. DeepMind is being repositioned as a cash engine, not an advanced research lab.
Retail will read this as an AI bubble story. Smart money will read it as a market structure rotation. The real question is not whether DeepMind can catch up. It is what Anthropic’s TPU order says about its own runway. Buying 20% of Google’s silicon is a massive bet. It tells me that AI infrastructure is now a tradable market, and the buyer of compute is the best proxy for future frontier work. This is why I now track Anthropic’s infrastructure footprint the same way I track a DeFi treasury. If a treasury locks up long-term assets, the token gets a floor. If Anthropic locks that much compute, it is signaling six quarters of build runway. That is a stronger signal than any model benchmark.
The Takeaway
The lesson is uncomfortable. A company can be technically strong and strategically obsolete at the same time. Google will harvest revenue. Anthropic will build the next architecture. In a bear market, this is the kind of resource shift I care about. Reputation does not keep a position alive. Allocation does.
Calculate. Execute. Repeat. The next time someone pitches an AI trade, ask one question: who holds the compute? If your project rents its own edge to a competitor, the edge is gone. DeepMind did not die in a single quarter. It was reallocated in small, visible orders. Liquidity vanishes. Lessons remain.