Where liquidity hides, narrative finds its voice. And right now, the most interesting liquidity signal isn't in a token pool or a bond auction—it's in the human capital flows of the AI industry. The silence between the blockchain blocks is deafening these days, but the echo from the AI sector is impossible to ignore. We watch capital flows, but we should also watch talent flows. They are the same force wearing different masks. When a company quietly hires the man who shipped seven generations of Google's TPU, it's not a press release—it's a balance sheet statement.
In a bear market, survival matters more than gains. And for the largest players in tech, survival is now defined by control over their own compute destiny. The macro signal here is not about digital assets directly, but about the broader liquidity landscape—the cost of capital, the concentration of supply chains, and the strategic imperative to own the means of production. This is a story about Anthropic's strategic pivot, but it's also a story about how any company that depends on a single supplier for its most critical input is living on borrowed time. We are chasing ghosts in the algorithmic machine, but these ghosts have names and résumés.
Over the past several months, the signals have been building. The chatter in the infrastructure and venture capital circles I orbit in Bangkok and beyond has been a mix of skepticism and intrigue. I have watched, with a sense of déjà vu, as the AI industry begins to mirror the crypto industry's own evolution from dependence to self-sovereignty. We saw it in DeFi, where projects fought for total value locked. We see it now in AI, where the battle is for compute. The question is no longer who has the best model, but who controls the substrate upon which the models run. The illusion of control in a fluid world is that you can rent your core infrastructure. The reality is that you either own it, or you are a tenant forever.
Context: The Human Signal
Anthropic, the force behind the Claude family of models, has reportedly hired Amir Salek. This is not a name that will resonate with the general public, but in the world of high-performance computing, he is a legend. Salek was the Vice President and General Manager of Google's TPU business, involved in the release of the first seven generations of the Tensor Processing Unit. His move to Anthropic is not a mid-level management shift; it is a statement of intent. To understand why this matters, you have to map the current flow of capital and control in the AI sector.
Anthropic currently sources its compute from a multi-vendor supply chain: NVIDIA, Google Cloud, and Amazon AWS. On the surface, this looks like a healthy hedging strategy. But in practice, it's a passive position. They are renting the intelligence that powers their own intelligence. They are at the mercy of their suppliers' price lists, allocation policies, and product roadmaps. It's the same trap that crypto protocols fell into when they became dependent on centralized oracle networks. The appearance of optionality was just a mask for total dependency.
This is where Salek's background becomes more than just a resume bullet point. His expertise is not just in chip architecture; it's in the full stack—from the semiconductor design to the compiler, the software stack, and the datacenter deployment. This is the experience of taking a custom silicon idea and turning it into a globally scaled system. This is exactly what Anthropic needs if it wants to move from being a pure model company to an infrastructure-defining entity. We are seeing the beginning of a shift from "buying compute" to "defining compute."
The trend is not isolated. OpenAI is already moving forward with its Jalapeño project, a custom inference chip developed in partnership with Broadcom. Google has its TPU, which is the only custom AI chip at scale. AWS has Trainium and Inferentia. In this map, Anthropic is the last of the majors that does not have its own silicon. This move is not an option; it's a strategic necessity to stay in the game. It is a defensive move to avoid being caught in a structural vice between the model's competition and the hardware suppliers. The core insight is that the war for AI dominance is no longer just about the algorithmic intelligence but about the physical substrate that brings it to life.
Core: The Macro and Micro of Custom Silicon
The decision to hire an expert like Salek speaks to a specific technical route. Based on my analysis of the macro trends and the pressure on capital, Anthropic is likely not aiming to build a general-purpose GPU to compete with NVIDIA. That would be a fool's errand. Instead, the likely path is a custom ASIC (Application-Specific Integrated Circuit) tailored to the specific load of the Claude model family. This is the pragmatic, capital-efficient approach that we saw in the world of DeFi when protocols realized they couldn't be all things to all people; they had to specialize in their own lane. The core insight here is that the value is not in the silicon itself but in the integration of the silicon with the model architecture.
This is where my perspective as a macro watcher comes in. The goal is not just to reduce dependency but to optimize the unit economics of inference. When you run a model like Claude, the cost of inference is the critical bottleneck. If Anthropic can design a chip that is 30% more efficient for their specific model's attention mechanisms, long-context windows, or MoE (Mixture of Experts) architecture, they can either lower their API prices or increase their margins. This is the yield trap, but this time it's in the compute layer.
The data from the crypto world is clear: the cost of a transaction is the cost of the security. The cost of an API call is the cost of the compute. If you can control that cost, you control the narrative. I have observed this in my analysis of the NFT market, where floor prices were highly correlated with stablecoin liquidity cycles. Here, the API pricing is correlated with the cost of compute. The ability to shift that cost curve is a massive strategic weapon.
A key insight that is often missed in the mainstream coverage is the importance of the software stack. The real moat for TPU is not the hardware but the compiler and the software ecosystem around it. The Tensor Processing Unit's performance is deeply tied to the XLA compiler and the software that optimizes the graph for the hardware. If Anthropic is serious about this, they will also be building a custom compiler and an operator library, not just the chip. This is a long-term, complex engineering project that requires the kind of cross-disciplinary talent that Salek brings. It is the difference between chasing ghosts in the algorithmic machine and actually being the architect of that machine.
From an investor's perspective, this move alters the valuation narrative. Anthropic is no longer a pure software play; it is becoming a capital-intensive infrastructure company. This has a double-edged effect. On the one hand, it creates a more durable barrier to entry and a story of vertical integration that the market rewards. On the other hand, it increases the need for capital and lengthens the time to a return on equity. The market will need to see proof of performance, not just a story of intent. We have to watch the hiring signals, the partnership announcements with foundries like TSMC or design partners like Broadcom, and the details of their data center contracts.
Contrarian: The Unwinding of the Decoupling Myth
The contrarian view, the one that makes the market, is that this entire move is a massive capital sinkhole. There is a narrative in the AI sector that "custom silicon is a rich man's game" and that the barriers are so high that even a company with Anthropic's cash flow will struggle to see a return. I am more skeptical of the broad narrative of "decoupling." I have lived through the Terra collapse and learned that hidden leverage is the true systemic risk. In this context, the hidden leverage is the supply chain. A strategy that tries to decouple from the general market but fails to secure its own production is just a different form of fragility.
The risk is that Anthropic gets caught in the middle. They are too big to ignore the cost inefficiency of buying from NVIDIA, but they are too small to build a full-scale chip ecosystem like NVIDIA has. The risk is a capital trap. A multi-year project that consumes billions of dollars and a huge portion of the company's focus could derail their model innovation and commercial timeline. In a bear market, this is the biggest risk: spending dry powder on a low-liquid asset.
However, the opposite angle is even more fascinating. The move toward custom ASICs might not be about the actual silicon but about the negotiation power. By having a credible in-house team, Anthropic changes the dynamics of its negotiation with its suppliers. The threat of in-house production is a powerful lever to get better pricing and allocations from a supplier. It's the same dynamic we see in the DeFi sector when a large protocol threatens to fork a codebase to get a better deal from a security provider. The signal of the team itself is a strategic tool, even if the chip never reaches mass production.
This is where the "decoupling" thesis fails. It's not about decoupling from the suppliers entirely; it's about re-coupling on their own terms. The strategy is to build the optionality. I am reminded of how the crypto market responds to the M2 money supply. The lag is not the disappearance of the signal, but the lag in the response. This is a 14-day lag that I've seen in the NFT market, and we might be seeing a similar lag here. The market hasn't priced in the full implications of this hiring because it doesn't recognize the long-term macro shift: the AI industry is moving from a rent-based model to an asset-based model.
The timing of this move is also crucial. We are in a period where the cost of capital is high, and the market is punishing unprofitable growth. But the macro liquidity cycle is starting to turn. The promise of cheaper compute is the key to the next wave of adoption. If they can get this right, they'll not just be a model vendor; they will be a key part of the foundation layer.
Takeaway: Tracing the Echo of a Viral Moment
The hiring of Amir Salek is a signal that the AI world is evolving into a system-level competition. It is not just about the model anymore. It is about the entire pipeline: model, chip, compiler, data center, and energy. The real value is in the integration of these layers. We are watching a shift from a world of virtual machines to a world of integrated platforms. The question is no longer whether Anthropic will build a chip, but what kind of chip they will build, and for what specific purpose. Will it be a high-performance training chip, or a cost-efficient inference engine? The answer will define the next decade of the AI industry. Reading the silence between the blocks, the next movement is already being coded. The question is whether the rest of the market can read the wiring before the ghosts disappear into the silicon.