The most revealing thing about a trade show is who is not allowed to stand on the floor. In late 2024, a single sentence began circulating through the European AI infrastructure circuit: Nvidia had reportedly forbidden neoclouds โ the newer, leaner GPU clouds like CoreWeave, Lambda, Crusoe, Nebius, and Together AI โ from sponsoring or exhibiting at GTC Berlin if they also deployed competitor silicon. The report carried no primary source. No timestamp, no quoted clause, no named executive. Just a whisper dressed as news. And yet the whisper is more analytically interesting than most confirmed press releases, because of what it implies about the shape of the moat Nvidia has spent a decade building.
I have spent the better part of two years auditing GPU supply chains, sitting in procurement calls with cloud resellers, and tracking the quiet migration patterns of developers who leave CUDA not because they want to, but because a customer demanded a second source. What I have learned is that the story of AI compute is almost never told honestly by either side. Nvidia tells a story of performance superiority, which is true at the frontier and increasingly irrelevant at the margin. The challengers tell a story of open ecosystems, which is true in spirit and painfully false in maturity. The neoclouds tell a story of neutrality, which is the most fragile fiction of all, because every one of them lives or dies by an allocation letter from Santa Clara.
So let us be honest about what we can and cannot know here. The factual record is nearly empty. One claim: Nvidia reportedly restricted neocloud participation at GTC Berlin. Everything else in this piece is structural inference, drawn from known behavior patterns, from the documented history of GTC sponsorship terms, and from the competitive logic that governs a market where one vendor holds an estimated eighty-plus percent share of AI accelerators. I will flag where I am extrapolating. I will not pretend a rumor is a smoking gun. But I will argue, with some force, that the rumor's plausibility is itself the finding โ because the fact that this story spread and was believed without evidence tells you exactly how the market expects Nvidia to behave.
GTC is not, in any meaningful sense, a trade show. It is an ecosystem loyalty ceremony. The keynote establishes the roadmap, the sponsorship tiers establish hierarchy, and the exhibition floor establishes which companies have been granted the right to describe themselves as members of the CUDA economy. A booth at GTC is a credential, not a marketing expense. Which is precisely why controlling it is so cheap. Nvidia does not need to threaten anyone. It does not need to withhold a single H100. It only needs to make the price of admission explicit, and every neocloud in the room will do the math themselves. A sponsorship package might run into the low six figures. A multi-year GPU allocation runs into the hundreds of millions, sometimes billions, in revenue and enterprise value. Nobody in this industry is going to trade a supply relationship for a floor plan. The exclusion enforces itself.
That self-enforcing quality is what makes this case so instructive. It is a textbook example of what competition lawyers call exclusive dealing, executed with the lightest possible touch. And it arrives at precisely the moment when the regulatory temperature around Nvidia has been rising on both sides of the Atlantic โ with the U.S. Department of Justice reportedly issuing subpoenas and the European Commission reportedly circulating market questionnaires to ecosystem participants. The timing of the report relative to that regulatory pressure is not coincidental. It is a marker of a company beginning to defend a position it previously did not need to defend.
Here is where I depart from the consensus reading of the neocloud wars. The prevailing narrative says Nvidia is fighting AMD. I think that is wrong, or at least incomplete. AMD's MI300X is a genuinely capable part; its HBM capacity draws at that dimension often exceeds what Nvidia ships. The shortfall is ROCm, not silicon. Microsoft, Meta, and Oracle have all made deployment commitments, and those commitments are real, but they are hedges rather than migrations. If the battle were purely about hardware, Nvidia would be competing on price and performance, as it has done for a decade. Instead, it appears to be competing on access. When the market leader starts defending the distribution channel rather than the product, it is telling you where it believes its own vulnerability lies.
The competitor Nvidia should actually fear is not AMD. It is Google. A TPU is not a chip you buy; it is a chip you access, bundled with a cloud platform, a customer base, and an integration story that requires no ROCm port and no new ops team. Anthropic's use of TPU infrastructure is the template: the customer never leaves the Google ecosystem, never fights a software stack, never encounters a sponsorship clause. AWS Trainium follows the same logic. These are alternatives that bypass the neocloud channel entirely, and therefore bypass the precise choke point Nvidia appears to be defending. AMD, by contrast, still needs third-party channels to reach customers who cannot build their own clusters โ and those channels are the ones being pinched.
This is the contrarian core of the whole affair. The conventional interpretation is that the exclusion proves Nvidia's power. I read it as evidence of a specific, bounded anxiety. CUDA's moat has never been raw performance; it has been migration cost โ the accumulated weight of tooling, libraries, debugging muscle memory, and the grim institutional certainty that porting a working training run is how projects die. If that migration cost were as deep and self-sustaining as the ecosystem claims, you would not need to administratively reinforce it. A moat that must be patched with sponsorship contracts is a moat whose designers suspect it is shallower than advertised. That is not a fatal wound. It is a signal, and signals precede revaluations.
Now consider the neoclouds themselves, the supposed victims. The victim framing is emotionally satisfying and analytically lazy. These companies are not neutral infrastructure providers; they are the marginal buyers whose growth most directly drives Nvidia's revenue upside, and several of them sit on the receiving end of Nvidia's own investment capital and supply commitments. Their customer bases are young, their sales cycles are short, and their switching costs are comparatively low โ which makes them the most contestable channel in the entire market. From Nvidia's perspective, they are not victims. They are the soft flank, and the soft flank gets fortified first. When a neocloud weighs a sponsorship package against an allocation relationship, there is no decision to make. The exclusion costs Nvidia nothing because the industry performs the punishment voluntarily.
The most counterintuitive consequence is also the most likely one. This episode will not trigger a mass neocloud defection to AMD. It will trigger the reverse. Smaller operators will move closer to Nvidia, deepening dependence to protect supply, producing the paradox that the more aggressively the channel is locked, the more concentrated it becomes in the short run โ even as the long-run political will for diversification hardens in every boardroom in the industry. The operators with genuine optionality are the hyperscalers: Microsoft, Meta, Amazon, Google, and OpenAI all have in-house silicon programs or dual-sourcing capability. They will read this report, file it under vendor risk, and accelerate their internal roadmaps. That is where the real cost lands, and it lands years out.
Europe deserves its own paragraph here, because GTC Berlin is not an arbitrary venue. Europe is simultaneously AMD's most established territory and the world's most aggressive competition-law jurisdiction. Executing an exclusion policy in that theater sends a message that is simultaneously commercial and legal, and it lands directly into a "sovereign AI" policy conversation that has been building for two years. The European Commission's TFEU Article 102 framework prohibits abuse of dominance including exclusive dealing and refusals to deal; the U.S. Sherman Act Section 2 prohibition on monopolization operates on similar ground. Nvidia's defenses would likely rest on the mundane observation that sponsorship rules for a privately organized conference are ordinary business practice โ the company is under no obligation to give floor space to competitors at its own event. That defense has real force at the "showcasing" end and weakens considerably if the clause actually reaches into what customers deploy in their own facilities. The legal character of this policy swings entirely on a distinction the reporting does not preserve: banning display is event management; banning deployment is market foreclosure. I have seen no text that settles which it is, and I will not pretend otherwise.
Which brings us to the honest accounting of gaps. I do not know which neoclouds were actually excluded. I do not know whether Nvidia's own portfolio companies were carved out. I do not know whether the restriction was accompanied by any explicit supply signal, whether it applies to GTC events globally or only Berlin, or whether the sponsorship term is longstanding boilerplate that simply received unusual enforcement this cycle. I do not know whether any European cloud operator outside the neocloud taxonomy was affected. Nvidia has not issued a public clarification that I can verify. Anyone telling you they have the full picture is selling you something, and given the industry we work in, the thing they are selling is probably a token allocation.
What I do know is that the episode maps cleanly onto a pattern I have watched play out three times now, in three different sectors of this industry: a dominant player, sensing that its technical lead is no longer self-evident, reaches for the softer levers โ distribution, credentialing, ecosystem access. It works. It always works, briefly. And it always leaves behind a residue of institutional memory in the customer base, a private note in every procurement file that reads: this supplier will use its position against us when it feels the pressure. That note compounds. It does not expire. It gets consulted the next time a second source is evaluated, and the time after that. Nvidia is not destroying its position. It is mortgaging a portion of it for a few quarters of channel stability, and the mortgage is priced in a currency โ trust โ that does not show up on the balance sheet until it is called.
The year ahead will tell us which reading is correct. If the coming quarters show stronger-than-expected neocloud consolidation around Nvidia, tighter supply allocation, and no meaningful acceleration in AMD or TPU adoption by mid-tier clouds, then the moat still holds and this exclusion was cheap insurance. But if we start seeing certifications for alternative silicon multiply inside the very channels that were locked out โ if the same neoclouds quietly announce multi-accelerator offerings to their enterprise customers while keeping their Nvidia allocations intact โ then we will know this clause was not a display of strength but a symptom. Watch the booths that were empty in Berlin. Some of them will tell you, six months from now, more about the future of compute than any keynote ever did. The industry is being asked, once again, whether it trusts the code or the hand that distributes it โ and for the first time in a decade, the answer is not obvious.