"Whoever wins in artificial intelligence will win the future."
I read it on September 3, in a wire brief no longer than a breath. One sentence from a podium. No annex, no named agency, no thresholds, no timeline. The same breath assured listeners that safeguards could exist, but warned that the conversation should not be "overly negative."
It should have scrolled past me. Instead it sat in my chest for a week.
What unsettled me was not the claim. It was the grammar. Win. A future is not a prize you take from someone else; it is a set of conditions you either build or fail to build. The sentence converts a shared timeline into a scoreboard, and once a timeline becomes a scoreboard, the only question left is who is permitted to keep score.
I have spent twelve years inside an industry that made this exact grammatical mistake and paid for it in ways that never appear on a chart. In 2017, I sat in a Manila hackathon room as the only woman among fifty engineers, listening to men say we would "win" decentralization. From the ashes of 2022, we planted seeds for 2030. Most of those seeds never germinated, and the ones that did were not the loud ones.
The architecture this sentence landed inside
Before a statement like this can mean anything, it has to be placed in the structure it falls into.
For three years, AI governance in the West ran on a premise of responsible pacing. The October 2023 executive order set a reporting threshold at 10^26 floating-point operations — roughly the compute scale of a frontier training run — and paired it with red-team disclosure and weight-security obligations. The Bletchley Declaration followed weeks later, with twenty-eight countries and the EU agreeing that frontier risk was a shared problem. Seoul added commitments in 2024. The EU AI Act phased in general-purpose model obligations through August 2025 and reserved its heaviest duties for high-risk systems later this decade.
Then the framing changed. The 2023 order was revoked in January 2025. What replaced it was less a framework than an attitude, and the attitude is now being spoken aloud: speed is the policy.
Here is the detail worth holding onto. Over the same period, the United States governed compute at the border far more aggressively than it ever governed models domestically. The export-control rounds of October 2022 and October 2023 built a perimeter around advanced accelerators, tuning thresholds by total processing performance and performance density specifically to catch chips designed to slip underneath them. A country reluctant to impose mandatory safety evaluation on its own labs was entirely comfortable deciding which other countries could buy which silicon.
That asymmetry is the tell. The instinct is not to constrain capability. It is to control the pipes capability flows through.
And crypto readers already know this instinct by name. We met it as a sanctions designation on a mixing protocol in 2022, and then as a set of developer prosecutions that turned on a very specific premise: not that the code was dangerous in itself, but that it could not be supervised. A system that cannot be monitored cannot be permitted. That sentence is not an AI policy. It is a governance reflex, and it does not care what the system is.
"Don't be overly negative," placed next to "safeguards," is not a red line. It is a disclaimer.
The technical work the wire brief skipped
The most common claim in our corner of the internet is that decentralized compute will route around the chokepoint. I think that claim is half true, and knowing which half matters more than the slogan.
Frontier pretraining is latency-bound, not throughput-bound. A run at the scale of 10^26 operations is not a pile of independent jobs; it is one synchronous computation. DeepSeek-V3 processed 14.8 trillion tokens across a cluster of roughly two thousand accelerators, and every gradient step required collective communication across all of them. That communication ran over NVLink-class interconnect — on the order of 900 gigabytes per second of aggregate bandwidth per accelerator — or its substitutes. Now imagine the same all-reduce spread across consumer GPUs in four countries over public internet links two orders of magnitude slower. The job does not slow down gracefully. It stops. Bandwidth does not average across a network the way price does.
So the honest split is this: decentralized networks are structurally credible for inference, LoRA-scale fine-tuning, batch rendering, and long-tail workloads where the unit of work is independent. They are not credible, at any near-term hardware generation, for frontier pretraining. Anyone selling you the latter is selling the story, not the silicon.
I say that as someone who has spent two years tracking DePIN compute tokens and wanted the answer to come out the other way. What I kept finding in the dashboards was supply subsidized by emissions, and utilization measured against a denominator that shrinks the moment the subsidy tapers. An economic model held up by its own token issuance is not a cost advantage; it is a deferred bill, and the bill comes due the day emissions halve.
Then the harder problem, the one that never makes it into the deck.
How do you know a remote GPU actually ran your computation? Proof-of-learning schemes that reproduce training traces remain impractical at frontier scale. So the practical answer today is hardware attestation — confidential computing modes on data-center accelerators, secure enclaves, vendor-run attestation services. That is a real solution. It is also a solution where the root of trust belongs to a chip vendor, and where revocation happens at a layer you cannot fork. You decentralized the scheduler and moved the chokepoint one floor down, into the microcode.
Zero-knowledge proofs of inference are the genuinely principled answer, and I want to be precise about where they stand. For small circuits, the overhead is manageable. For large models, published estimates run from three to six orders of magnitude above the cost of simply running the model. That is not a rounding error. That is a research program.
Which brings me to something I have not seen anyone connect, and I believe it is the most important structural fact in this essay. Rollup provers and AI proving markets are bidding for the same scarce resources: high-memory accelerators, proving farms, cryptographic engineering talent. Every rollup that commits to validity proofs is competing against a machine-learning market orders of magnitude larger. Post-Dencun blob space feels cheap right now, and blob demand is priced against a fee curve designed when the only consumer of data availability was rollups. When inference and proving draw on the same silicon, your Layer 2 gas fee stops being a Layer 2 decision. It becomes a line item in an AI capital-expenditure cycle, and it reprices on someone else's schedule.
The final place this lands is the one that should worry all of us more than model weights do.
In 2025, agent payments stopped being a thought experiment. x402 turned the dormant HTTP 402 status code into a settlement rail. ERC-8004 gave agents on-chain identity and reputation primitives. The result is an economy where machines transact at a frequency and granularity that makes per-transaction human review structurally impossible — not inconvenient, impossible.
Once human review is off the table, there are exactly two designs, and they cannot both win. You can build programmable permissioning: attested agent identities, allow-lists, reputation scores, revocation rights, compliance hooks. Or you can build permissionless settlement: bearer assets, no gatekeeper, no ability to un-send value. The first is a surveillance architecture with better branding. The second is the thing crypto was actually for.
I will tell you which one arrives by default, and it is not the second. Enterprise procurement requires an identifiable counterparty. Liability law requires someone to sue. Anti-money-laundering compliance requires the ability to freeze. None of those requirements are evil; they are simply incompatible with bearer settlement. So the default is a world of rated, revocable, identified machine agents transacting in a programmable unit whose permissions live one layer above the money.
The fight over central bank digital currencies and open money was never going to be argued on human wallets. That was too politically expensive. It arrives instead as agent identity, trust frameworks, and machine payment compliance — dressed in the language of safety, and adopted without a vote. If you are building in the agent stack right now, ask one question of your architecture: does it assume the right to revoke? If the answer is yes, you have already chosen. You chose before the policy did.
Where my own side is wrong
The reflexive crypto answer to everything above is "decentralize it." That is a slogan wearing an architecture costume. Our most vocal decentralization advocates are frequently capitalized by the very compute concentration they critique, and the "decentralized" networks we celebrate run on silicon from a single Taiwanese foundry, lithography from a single Dutch monopoly, and design from two American firms. We did not exit the chokepoint. We built a very elegant interface on top of it and called the interface sovereignty.
There is a second blind spot, and it is less comfortable to name. The decentralization movement has no credible answer to catastrophic misuse at the frontier, and pretending the only real risk is state overreach is its own kind of debt — the kind that compounds quietly while everyone is busy being right.
I learned the shape of that debt in 2017, when I spent weeks studying a whitepaper I wanted to believe. Decentralization as a shield is how bad actors ask for your trust, and it works, because we have been trained to hear the word and relax.
So here is the pragmatist's test I now apply to every protocol claiming to fight the chokepoint. Not "is it decentralized" — that question has become unfalsifiable. Instead: does this design remove a chokepoint, or does it relocate one to a layer where you can no longer see it? A scheduler in a smart contract that trusts a vendor attestation service has removed nothing. It has hidden something.
What to actually watch
That sentence from the podium will be quoted for years, and its power is that it reframes a commons as a contest. But contests need scorekeepers, and scorekeepers need records.
The open question of the next decade is not who wins artificial intelligence. It is who can prove what happened — who holds the receipts, verifiable by anyone, without asking permission from the party with an interest in the answer.
That is the contribution open networks can still make. Not to win the race. To keep the ledger.
