When a company that burns through billions in GPU credits hires the engineer behind seven generations of TPUs, it is not merely diversifying supply chains; it is rewriting the definition of sovereignty in the digital age. On March 3, 2026, Anthropic announced the appointment of Amir Salek, former Google vice president of TPU and Silicon, to lead its semiconductor strategy. The move comes as the AI model maker simultaneously discloses a multi-year, multi-cloud procurement agreement with NVIDIA, Google Cloud, and AWS—a triangulation of compute that reads less like a supply deal and more like a hedge against a future where no single vendor can be trusted with the keys to the kingdom.
For a cross-border payment researcher who has spent the last decade mapping the hidden costs of financial intermediation, the pattern is achingly familiar. Just as stablecoins emerged to bypass the correspondent banking system, AI firms are now trying to bypass the chip oligopoly. The difference is that stablecoins trade in liquidity, while chips trade in physics—and physics is far harder to fork.
Context: The Macro Liquidity Map of Compute
The global AI compute market is currently anchored by NVIDIA’s H100/B200, Google’s TPU v5, and Amazon’s Trainium/Inferentia. These are not just chips; they are the underlying infrastructure upon which the next generation of financial settlement, identity verification, and cross-border remittance systems will be built. As of early 2026, the total annual spend on AI training and inference exceeds $120 billion, a figure that dwarfs the total market cap of all cryptocurrencies excluding Bitcoin. The liquidity of compute—the ability to access it at scale, on demand, and at a predictable price—has become the new capital adequacy ratio for any company that touches digital assets.
Anthropic’s current position is precarious. It spends approximately $1.5 billion per year on cloud compute, with the majority flowing to NVIDIA-based instances on AWS and Google Cloud. Its Claude models are among the most compute-intensive in the world, particularly for long-context reasoning and tool-use tasks. Yet it has no control over the silicon that powers its own products. The hiring of Salek, who oversaw the TPU program from its first generation through the v7, signals a clear intent to move from a pure rental model to a hybrid architecture where custom silicon handles the most critical workloads.
Core: The Crypto-Infrastructure Lens
From a macro watcher’s perspective, this move has three direct implications for the blockchain ecosystem.
First, the competition for fab capacity will intensify. TSMC’s 3nm and 2nm nodes are already oversubscribed, with Apple, NVIDIA, and AMD consuming the majority of advanced wafer starts. If Anthropic and OpenAI (via its Jalapeno project) both bring custom ASICs to production, they will compete for the same physical slices of silicon that might otherwise go to Bitcoin mining ASICs, Ethernet switches for validator networks, or zk-proof accelerators for layer-2 rollups. The marginal cost of compute for every blockchain from Ethereum to Solana will rise if AI chips absorb capacity. I have seen this dynamic before: in 2021, when GPU shortages made Ethereum mining unprofitable for small miners, the same scarcity drove the NFT market’s carbon footprint to absurd heights. Today, the scarcity is about physical wafer area, not just hash rate.
Second, the rise of custom AI chips undermines the narrative of decentralized compute networks like Bittensor, Render Network, and Akash. These platforms promise to commoditize AI inference by pooling consumer-grade GPUs from around the world. But if the most efficient inference hardware is proprietary, single-tenant, and tightly coupled with a specific model architecture, the utility of a decentralized network drops. The hollow resonance of digital sovereignty in hardware silos mirrors the centralization crypto was built to resist. Bittensor’s subnet validators, for example, rely on open-source ML frameworks that run on commodity GPUs. If Anthropic’s chip is optimized for Claude’s MoE architecture with a custom instruction set, that chip will never be available to the open network. The result is a two-tier compute market: one for the incumbents, one for the rest.
Third, the stablecoin and cross-border payment use case is directly affected by inference cost. Claude’s API pricing, which currently sits at $15 per million tokens for long-context models, is a function of compute cost. If Anthropic can reduce inference cost by 40% through custom silicon, it will pressure other AI providers to follow suit. That, in turn, will accelerate the deployment of AI agents for AML screening, fraud detection, and remittance routing—all of which are currently bottlenecked by the cost of running large models. Based on my audit of cross-border payment rails, I have seen that the cost of compute is the new friction, replacing the old friction of correspondent bank fees. Every percentage point reduction in inference cost directly expands the addressable market for on-chain compliance tools.
Contrarian: The Decoupling Thesis
The conventional wisdom is that Anthropic’s chip push is a threat to NVIDIA and a validation of the “AI monopoly” narrative. But the contrarian view—the one that aligns with my structural skepticism of decentralization—is that the move is actually a desperate attempt to escape a trap of the incumbents’ own making. Anthropic and OpenAI are both deeply dependent on cloud providers that are also their competitors. Google owns TPU and also develops Gemini. Amazon has its own AI models through Bedrock. NVIDIA is building its own full-stack AI platform. The AI firms are not becoming more powerful; they are becoming more vulnerable to supply chain blackmail.
In crypto terms, this is the equivalent of a DeFi protocol that relies on a single oracle. The protocol can be technically sound, but if the oracle is corrupted or becomes unavailable, the protocol fails. Anthropic is trying to become its own oracle—but in doing so, it is taking on the risk of being a hardware company. Regulation lags, capital moves, but hardware moves even slower. The lead time for a custom ASIC is 18 to 24 months, and the cost is in the hundreds of millions of dollars. The risk that Anthropic’s chip will be obsolete before it reaches production is real, especially given the pace of NVIDIA’s architecture roadmap.
Moreover, the chip project will consume resources that could otherwise be used for model training, safety research, or enterprise sales. If Anthropic’s investors (including Google and Amazon) see the chip project as a distraction, they may pull support. Liquidity evaporates when trust fractures, and trust in AI companies is already fragile after the 2025 model collapse scandals. The contrarian bet is that Anthropic will either fail to deliver a competitive chip or will succeed only to find that the market no longer needs it because decentralized inference networks have solved the cost problem through algorithmic innovation rather than hardware.
Takeaway: Cycle Positioning
The next cycle in crypto will be defined not by which model wins, but by who owns the silicon that runs it. My analysis of the hiring of Amir Salek leads me to a single forward-looking judgment: Anthropic’s chip ambition is a signal that the compute layer of the AI stack is transitioning from a commodity to a strategic asset, and that transition will create both opportunities and risks for blockchain infrastructure. The opportunity is that cheap, proprietary inference chips will drive demand for tokenized settlement of AI services—if those chips are paired with on-chain micro-payment rails. The risk is that the same chips will become the new choke points, centralizing the very infrastructure that crypto seeks to unbundle.
In the Alpine quiet of Geneva, I watch this unfold with a sense of déjà vu. The same pattern that played out in banking—centralization, then regulatory arbitrage, then fragmentation, then re-centralization—is now playing out in compute. The difference is that this time, the tools of resistance are not just code, but also the physical limits of silicon. The question is not whether Anthropic will build a chip, but whether the broader ecosystem will build the economic layers that make that chip accessible to all, rather than another walled garden. The hollow resonance of digital sovereignty in hardware is a warning, not a prophecy. I intend to listen.