Consider the moment when a founder I know in Tallinn opens her laptop to an email she has been dreading. Her cloud provider is cutting her reserved GPU allocation by forty percent next quarter. The reason is not her credit score, not her usage patterns, not any failure of her own. The reason is a single sentence that has become the defining refrain of this bull market: demand exceeds supply. She is building an open-source agent on top of decentralized identity rails. By any reasonable measure, she is exactly the kind of builder the industry claims it wants. And the infrastructure she depends on has quietly decided she is not a priority. This is not a story about one startup. It is a story about who gets to compute in the age of artificial intelligence โ and what the blockchain community, which claims to care about access, should do about it.
The sentence itself came from AMD chief executive Lisa Su, who told the market that demand for her company's products exceeds its ability to supply them. It arrived without a timestamp, without a named product line, without a financial figure. That absence is itself the story. When the leader of the company that designs the second-most-important AI accelerators on earth says demand exceeds supply, she is not describing a manufacturing miracle. She is describing a choke point.
And that choke point is not where most people assume. AMD is fabless. It owns no fabrication plants. Its most advanced wafers โ three-nanometer and four-nanometer โ come from a single supplier, TSMC, in a single geography. Its Instinct MI300X and MI325X accelerators are assembled from chiplets bonded using TSMC's advanced packaging, the same CoWoS queue in which NVIDIA stands. Its memory comes from an oligopoly of three firms โ SK Hynix, Samsung, and Micron โ building HBM3E.
So when Su says supply cannot meet demand, she is telling us something structural. The entire AI economy is bottlenecked at two shared chokepoints: advanced packaging capacity and high-bandwidth memory. No amount of architectural brilliance can route around physics. AMD's hardware has essentially closed the gap with NVIDIA โ the two now ship on similar nodes, within roughly half a year of each other. But the software ecosystem, ROCm against CUDA, still trails by three to five years. And the supply chain beneath both is a single point of failure dressed in a hundred billion dollars of optimism.
There is a detail worth lingering on. The statement was positive in tone โ 'demand exceeds supply' is the language of strength, not weakness. But a positive sentence is not the same as a complete one. It omits the export controls that have severed AMD's access to roughly a fifth of its addressable market. It omits the fact that the same hyperscalers buying its accelerators are also building competing silicon in-house. It omits the software gap that decides whether developers actually adopt the hardware. In a bull market, we read the optimistic half of every sentence and treat the rest as noise. That habit is expensive.
Here is the synthesis that should concern anyone who believes in decentralization. Artificial intelligence is being concentrated into fewer hands than any technology in modern memory. One foundry packages the accelerators. One company designs most of them. Three firms make the memory. A handful of hyperscalers โ Microsoft, Meta, Oracle, xAI โ buy the output. The result is a supply chain that looks less like a market and more like a chokepoint diagram. This is precisely the centralization failure mode that the blockchain community was founded to resist.
I have audited this pattern before. In 2017, during the height of the ICO boom, I read more than fifty whitepapers and found only twelve with viable economic models. The lesson I took from that filtering exercise, which I later distilled into a manifesto on the human layer of blockchain, was simple: technology serves human trust; it does not replace it. The same lesson applies now. A token cannot manufacture a wafer. A governance vote cannot conjure a packaging line. The physical world sets the terms, and the token economy negotiates at the margins.
That is why the decentralized compute sector โ the so-called DePIN networks that promise to democratize AI infrastructure โ deserves both our attention and our scrutiny. The pitch is compelling. Instead of routing every inference request through a handful of hyperscalers, distribute it across a heterogeneous mesh of GPUs: idle gaming rigs, underused data-center capacity, edge devices. Pay the operators in tokens. Verify the work with trusted execution environments or zero-knowledge proofs. Make compute fungible across trust boundaries.
When compute is verifiable, it becomes transferable. And when it becomes transferable, trust stops being a matter of corporate reputation and becomes a matter of cryptographic proof. That is the genuinely radical promise. Trust is the only currency that matters, and verifiable compute is the first honest exchange rate this industry has ever offered.
But let me be precise about what these networks can and cannot do. Based on my audit experience, I apply a single test to every decentralized compute claim: where does the silicon come from? If the answer is 'we aggregate existing supply,' then the network is a broker, not a builder. Brokers are useful. Brokers are not bottlenecks. The actual bottleneck โ CoWoS packaging and HBM โ is owned by TSMC and a Korean-American memory cartel, and no DePIN token has yet bought its way into that room. Advanced packaging capacity expands on a twelve-to-eighteen-month cycle. HBM fabs take years and billions. You cannot govern your way to a shorter lead time.
So the honest framing is this: decentralized compute is a distribution-layer innovation, not a production-layer one. It reshuffles who gets access to the marginal GPU-hour. It does not create the GPU-hour. That is still an enormously valuable contribution โ because the current allocation mechanism is failing my founder in Tallinn, and it is failing thousands of builders like her. But we should stop pretending that token incentives solve a physics problem.
What they can solve is the allocation problem. Consider what happens when a hyperscaler decides a small team is not worth a reserved allocation. The team has no recourse. There is no appeal, no transparent queue, no way to prove its workload is legitimate. In a decentralized market, the same team competes on price and verifiable reputation rather than on relationship and scale. That is a genuine improvement in fairness, and fairness is what makes an ecosystem resilient.
There is a second, subtler benefit. Geographic concentration is a risk that the market systematically underprices. When the supply chain runs through one island, a single geopolitical event โ a blockade, an earthquake, an export-control regime โ can cascade through the entire AI economy. Decentralized compute, by definition, spreads inference across jurisdictions. It does not eliminate the risk at the fab layer, but it does reduce the risk at the deployment layer. Redundancy is not glamorous. It is also not optional.
Here is the double-edged nature of scarcity for a challenger. NVIDIA, holding something like eighty to ninety percent of the AI accelerator market, can afford to be allocation-constrained. Its customers have nowhere else to go. AMD, holding perhaps five to ten percent, cannot. If a hyperscaler cannot get MI300s on schedule, it does not wait โ it buys more Blackwell, or it accelerates its own silicon program. Google has TPU. Amazon has Trainium. Meta has MTIA. Every quarter of delay is a quarter in which a customer learns it can live without you. Scarcity validates demand, but it also punishes the supplier who cannot satisfy it. For the leader, scarcity is leverage. For the challenger, it is a countdown.
The customer concentration makes this sharper. AMD's AI revenue leans on a small set of buyers, and those same buyers are its most credible future competitors. This is a structural trap that no amount of chip design solves. The hyperscalers are simultaneously the demand signal and the threat to the demand signal. When I advise founders in the Estonian Web3 community, I tell them to map their customers' incentives before they map their own roadmap. The same discipline applies here: the companies queuing for your accelerators are quietly funding the teams that will replace them.
In 2025 I helped convene a small research group โ the Human-Centric AI Alliance โ to study how decentralized identity could protect privacy in the age of large language models. We reviewed twenty academic papers and ten pilot projects, and one conclusion stood out. The hardest problem in decentralized compute is not supply. It is proving that a computation happened correctly without trusting the machine that ran it. Trusted execution environments help, but they trust the chip vendor. Zero-knowledge proofs are pure but expensive. Verifiable human interaction โ the framework we proposed โ matters precisely because AI systems increasingly mediate between people and infrastructure. If we cannot verify who is asking and what was computed, decentralization becomes a slogan rather than a safeguard.
Now the part that the bull market does not want to hear. The decentralized compute sector is repeating the mistake that Layer2 networks made. There are dozens of networks now, all chasing the same small pool of GPU supply. This is not scaling. It is slicing an already-scarce resource into ever-thinner fragments. Each network launches its own token, its own incentive program, its own liquidity mining campaign, and each one competes for the same operators. The result is fragmentation dressed as competition โ a hundred marketplaces for a commodity that fewer than a dozen suppliers actually produce.
The governance story is equally uncomfortable. Every one of these networks advertises decentralization while retaining an upgrade key held by a multisig. Code binds, but people break or build. The smart contract is immutable right up until the moment three of five signers decide it is not. I have watched this pattern across DAOs for years: the whitepaper promises community sovereignty, and the reality is a foundation with a treasury and a privileged set of administrators. Decentralization becomes a compliance shield rather than a governance model โ the team wallets and foundation holdings are traceable, but the marketing language is not.
None of this means the sector is doomed. It means the sector is young, and youth is not an excuse for dishonesty. The networks that will survive are the ones that publish their supply-chain provenance, disclose their key management, and compete on verifiable delivery rather than on narrative. Culture eats blockchain for breakfast โ and the culture of quiet centralization will eat the compute narrative from the inside if it is not addressed.
The test I apply is the same one I applied in 2017. Ask what is real. Ask what is borrowed. Ask who holds the key. A network that passes those three questions deserves your compute and your capital. A network that cannot answer them deserves neither, no matter how compelling its bull-market narrative.
So where does that leave us? AMD's sentence โ demand exceeds supply โ is not really about AMD. It is a mirror held up to an industry that has concentrated the most important technology of our generation into the fewest possible hands. The blockchain community has spent a decade arguing that concentration is fragile. Here is the proof, written in silicon and shipping delays. The opportunity is not to pretend that tokens can build fabs. It is to build the distribution and verification layer that makes the physical bottleneck survivable โ and to do it with the honesty that the last cycle lacked. We are building the future, together, but only if we are willing to say out loud what the future is actually made of.