The Reasoning Premium: A DeepMind Alum's Quiet Fundraise and the Compute Layer Crypto Still Cannot Own

On a grey Geneva morning in early 2026, an alert crossed my desk from a crypto vertical: Thore Graepel, described with the loose brushwork that has become standard in trade media as a "co-creator" of Google DeepMind, had raised funds for a new AI reasoning startup. The item contained six facts and no numbers. No company name. No round size. No valuation. No investor list. No product. Just a pedigree and two sentences of public-relations vapour about "redefining AI's role in complex problem-solving." I read it twice, not because it was informative, but because it was diagnostic. A crypto publication was covering an AI reasoning round with zero blockchain content, in the middle of a bear market where crypto's own liquidity is evaporating, and the only reason the item existed was that the word "DeepMind" still carries enough gravitational pull to make an empty story travel. The hollow resonance of that headline was instructive. It told me, more clearly than any honest data release could, where capital believes the next durable margin lives — and where it does not. And as someone who has spent nine years auditing cross-border settlement flows, DeFi mechanism design, and the compute markets that promise to underwrite both, I have a specific interest in what the answer implies for the chain. To understand why a name and a tagline were sufficient to raise money is to understand the single most important structural shift in machine learning since the transformer: the migration of compute from training time to inference time. For most of the last decade, the frontier of AI capability was defined by how much computation you could throw at pre-training — the colossal, one-time burn that turns raw text into a foundation model. The scaling law that governed this phase was brutal and simple: more parameters, more data, more GPUs, better model. That era is not over, but it has been joined, and increasingly overshadowed, by a second scaling law that operates at the moment of use. Test-time compute scaling — spending additional computation during inference, letting a model "think" through longer chains of reasoning, sample multiple candidate solutions, and verify them — has become the dominant lever for pushing capability on hard problems. When OpenAI released its o-series models, when DeepSeek published R1 as an open-weight reasoning model, and when Google folded "thinking" modes into Gemini, they were all announcing the same economic transition. The centre of gravity in AI compute had shifted from a concentrated, predictable, capex-heavy training cluster to a distributed, continuous, and far more expensive inference load. That shift is the reason Graepel's background matters more than his company's name. His reputation was built on reinforcement learning, game theory, multi-agent systems, and search — the intellectual lineage of AlphaGo, AlphaZero, and Monte Carlo Tree Search. Reasoning and search are the same family. When you allow a model to explore a tree of possible thoughts and evaluate them, you are, quite literally, running a search algorithm inside a language model. The most credible technical bet in reasoning right now is the synthesis of reinforcement-learning post-training with structured search and verification. Very few people on earth have deeper intuition for that synthesis than someone who spent years teaching machines to play Go. So the fundraise is not nothing. It is a signal that the capital markets have correctly identified reasoning as the axis of competition for the coming cycle. But it is precisely at this point that the crypto reader should feel a specific kind of discomfort, because the story the blockchain industry wants to tell about itself and the story this fundraise actually tells are not the same story. For three years, the dominant thesis in crypto's infrastructure sector has been that AI compute — and especially inference — would eventually be commoditised, decentralised, and routed through permissionless markets. DePIN, decentralised physical infrastructure networks, was supposed to turn idle GPUs into a global inference fabric. The pitch was seductive: aggregating stranded compute, rewarding suppliers with tokens, and undercutting the hyperscalers on price. It mapped neatly onto the populist romance of decentralisation, and it attracted real capital. And yet here we are, watching the most important compute transition in a decade, and the money is flowing to a centralised startup founded by a corporate alumnus, using what will almost certainly be hyperscaler infrastructure, financed by venture capital that has no interest whatsoever in token distribution. The reasoning boom, the very trend that DePIN most needed to capture in order to justify its own valuation, is consolidating compute rather than dispersing it — because reasoning workloads have economic properties that are structurally hostile to the decentralised model. I want to be precise about why, because I have written the optimistic version of this argument myself, and I have had to walk it back. Based on my audit experience, the decisive difference is verification cost. A decentralised network of anonymous GPU suppliers can serve a deterministic workload — rendering, hashing, or a simple matrix multiply — because the output is cheap to verify and trivial to compare across independent nodes. You sample the work, you check it, you slash dishonest suppliers. The economics of redundancy are tolerable when verification is cheap. But a reasoning model's output is not a hash. It is a long, stochastic chain of intermediate tokens whose quality cannot be checked without re-running the very computation you were trying to distribute. Trustless verification of inference is possible in principle — zero-knowledge proofs of execution, optimistic challenge games, redundant computation with consensus — but each mechanism imposes a cost that scales with the very thing that makes reasoning valuable: its length and its ambiguity. This is the trap the Decentralised Physical Infrastructure narrative keeps walking into. It assumes all compute is fungible. It is not. Pre-training can be distributed. Fine-tuning can be distributed. But the frontier of value right now sits in long-horizon reasoning, and long-horizon reasoning is expensive to verify and therefore expensive to trustlessly decentralise. The projects that pretend otherwise are, in effect, subsidising their own metrics — running token emissions to pay suppliers to serve workloads whose verification they quietly centralise. That is the same pattern I documented in DeFi's liquidity mining era, and it produces the same result. Stop the subsidy and the suppliers vanish, because the underlying unit economics never closed. I remember the 2020 DeFi Summer with a clarity that still carries a faint bitterness. I spent that season inside Curve Finance's mechanism design, pulling apart more than five thousand pool transactions to understand how stablecoin pegs held. What I found was elegant and, in a specific sense, hollow: the efficiency was real, but it was propped up by opaque oracle dependencies and governance capture that the "decentralised" label papered over. The same cognitive dissonance is available to anyone who reads a DePIN token thesis carefully today. The aggregation of idle compute sounds novel. The aggregation of idle compute into a market whose hardest workload cannot be trustlessly verified is a business model in search of a problem it can actually solve. The second problem is liquidity and cost structure. A reasoning model does not merely cost more per query than a conventional language model — it costs dramatically more, because the "thinking" is the product. A single hard query can consume ten to a hundred times the tokens of an ordinary conversation. That is wonderful for anyone selling compute and catastrophic for anyone selling a margin. It means reasoning startups have a voracious appetite for continuous inference capacity, and it means their gross margins are hostage to the price of GPUs and the efficiency of their serving stack. In that environment, the buyer wants one thing above all: reliability. They want a contract, a service-level agreement, a vendor who will not vanish. They do not want to route their frontier workload through an anonymous marketplace with probabilistic uptime and punitive slashing conditions, no matter how attractive the token rewards look on a dashboard. This is where my second area of hard-won experience becomes relevant, and where the stablecoin analogy bites. I spent years in Geneva auditing the old correspondent-banking rails — the SWIFT messaging layer that still moves the world's cross-border money — and I interviewed forty migrant workers in Zurich about the true cost of sending money home. Thirty-five per cent of their transfers were being eaten by intermediary fees layered invisibly across the chain. Blockchain was supposed to fix that. It half did, and the half it fixed was the half that got regulated first. The lessons that survived are worth transposing to compute. The winner in a market defined by trust deficits is rarely the most decentralised participant. It is the participant that buys legitimacy by becoming the regulator's partner. That is exactly what PayPal did with PYUSD. It did not fight the regulatory perimeter; it embedded itself inside it. The rational strategy for any infrastructure provider facing existential compliance risk is to trade a measure of decentralisation for a measure of certainty. The reasoning startup that Graepel is building will almost certainly make that trade, and it will make it deliberately. It will use hyperscaler infrastructure. It will accept enterprise contracts. It will court procurement departments and possibly a sovereign customer or two. And it will out-compete the token-based alternatives not because it is more efficient in the abstract, but because it is more trusted at the exact seam where trust is scarce. Now consider the governance dimension, because it is the one crypto builders most consistently ignore. There is a parallel between the AI safety debate and the DAO liability debate that almost nobody draws, and it illuminates both. When we talk about a reasoning model, we are talking about a system that can plan. Planning capacity is what makes reasoning valuable, and it is also what makes it dangerous — a model that can decompose a hard task into steps can decompose a harmful one too. Regulators understand this. The EU AI Act, which I had the strange privilege of stress-testing in a roundtable between European regulators and decentralised compute developers in Geneva, anticipates exactly this class of risk. What struck me most in that room was not the ambition of the framework but the verification gap at its centre: roughly seventy per cent of the training and evaluation data in the pipelines we examined could not be traced to a verifiable source. The provenance was missing. The regulators wanted accountability, and the technology could not supply it. This is where blockchain has a genuine, narrow, and largely unclaimed contribution to make. Not as a place to run inference, but as a place to anchor truth. Zero-knowledge proofs of data provenance, on-chain attestations of model lineage, tamper-evident audit trails — these are problems where cryptographic commitment has a real comparative advantage, because the workload is verification, not computation. This is the thesis I have been developing for the last year, and it reframes the entire AI-crypto convergence. Crypto's role in the reasoning era is not to decentralise the compute. It is to verify the claims. And it is worth being blunt: the industry keeps chasing the first and ignoring the second because the first is a bigger market and the second is a smaller, harder, less glamorous one that actually works. Let me return to the fundraise with this frame, because the omissions now read differently. The item gave us no company name, no round, no investor, no product. Earlier this year I would have treated those gaps as a failure of the reporting. Now I read them as a feature of the moment. When a category is hot enough, the pedigree alone clears the round. The "DeepMind alumni plus reasoning" combination has become a standard venture formula, and the formula does not require a product to function — it requires only a plausible founder and a plausible direction. We saw this movie with Inflection, with Adept, and with a dozen other well-capitalised teams that ended up absorbed through talent acquisitions before their products ever reached a market. The光环 — the aura of pedigree — is real capital, and like all capital it can be overvalued. A tool named after a founder's past finds it easy to raise money and hard to raise accountability. So let me state the honest limit of what I know, because the discipline of a macro watcher is to distinguish signal from noise. The direction is clear: reasoning is the frontier, test-time compute is the frontier spend, and inference capacity is becoming the scarce commodity of the cycle. The specific bet is unverifiable: I cannot tell you whether this particular startup is building a genuinely non-consensus search-and-reinforcement architecture or simply engineering a longer chain of thought on top of someone else's open base. The difference matters enormously for investors and hardly at all for the narrative, which is precisely why the narrative is dangerous. Here is the contrarian claim at the heart of this piece, and I want to hold it steadily. The reasoning boom does not need crypto. That is not a defect of crypto — it is a structural fact of the workload. Reasoning companies need compute density, low latency, verification of their own outputs, and enterprise-grade trust. Decentralised markets are weak on all four and strongest on none. The honest reading of the Graepel fundraise is not that it threatens DePIN; it is that it happens in a parallel universe where DePIN's thesis is simply irrelevant. The two markets are increasingly decoupled, and the sooner the blockchain industry stops trying to attach its tokens to every adjacent AI headline, the sooner it can focus on the one seam where it genuinely adds value — verifiable provenance in a world drowning in unverifiable synthetic data. That decoupling has an investment implication that few are pricing. If reasoning value accrues to centralised players, then the beneficiaries of the cycle are the inference-chip designers, the serving-stack optimisers, the hyperscalers, and the few toolchains that can make long-context inference efficient enough to be profitable. The crypto expression of this trend is not the token that rewards idle GPUs. It is, at most, the small set of protocols anchoring data provenance and the infrastructure tokens that quietly ride the same compute demand without claiming to replace it. The rest is liquid mining in a new costume, and it will evaporate the moment the emissions stop. I have watched this pattern long enough to recognise its emotional signature. In 2021, I tracked the energy consumption of Ethereum's proof-of-work network and calculated that minting ten thousand high-profile art pieces exceeded the annual carbon footprint of a hundred thousand Geneva households. The number was real, and it clashed with everything I wanted the technology to be. I stopped writing for two months. When I came back, I came back sober, and I built a beat around sustainability and survival metrics precisely because that is where the durable value hides. The reasoning cycle has a similar structure: a genuine technical revolution, wrapped in a financing narrative that is mostly aura. My job is not to be thrilled. My job is to be accurate. So what is the forward-looking question that should occupy anyone reading the Graepel story in a crypto context? It is not "which token captures AI reasoning." It is this: as the frontier migrates from training to inference, and as inference consolidates into a handful of certified, compliant, well-capitalised vendors, what remains for a permissionless market to do — and will the industry have the discipline to build there rather than keep chasing the compute it cannot trustlessly verify? The companies that answer that question honestly will outlast the cycle. The ones that keep selling decentralised inference as though verification were free will discover, once the emissions stop, exactly how hollow the resonance of their narrative always was.