Semiconductor research firm SemiAnalysis reports that Nvidia has refused to let a number of "neoclouds" (small and mid-sized cloud providers that specialize in renting out GPUs) attend the GTC Berlin conference, because those companies have added non-Nvidia chips such as AMD GPUs and Google TPUs to their platforms. SemiAnalysis says it has confirmed the matter but has not published a list of company names.
GTC Berlin runs from October 20 to 22, with sessions at Tempodrom and STATION-Berlin. Reports say the affected providers were notified only a few days before the opening. As of publication, Nvidia has not publicly responded, and everything below comes from the SemiAnalysis report and accounts relaying it, none of which can currently be verified.
Aimed only at the small players
According to SemiAnalysis, the pressure is "selective." Hyperscalers such as AWS and Microsoft Azure also buy AMD hardware and build their own chips, yet they have not faced similar treatment. The explanation given in the report is that their purchasing volume is large enough that Nvidia has little leverage over them.
Neoclouds are in a different position. Nvidia is their most important chip supplier and a shareholder in companies such as CoreWeave, and it has taken part in the financing arrangements of some providers in various ways. For a neocloud of limited size, where it stands in the queue for the latest batch of GPUs, and whether it can count on Nvidia's support for financing or a public listing, can decide whether the business survives.
The report says some providers have begun considering a change of supplier, and some executives are weighing whether to offer non-Nvidia chips "quietly" so as not to provoke Nvidia.
Three milestones over a few months
Set against the timeline, this did not come out of nowhere:
- June: SemiAnalysis published a set of discussions in which several neocloud executives privately voiced concern that adopting AMD GPUs or TPUs, or even just using non-Nvidia networking equipment, could cost them early supply allocations or Nvidia's backing for financing and IPOs.
- August: The report said many neoclouds were already ordering AMD Instinct and Nvidia GPUs at a ratio close to 1:1, and that at least two had deployed 10,000 GPUs from each vendor.
- October: The fear became reality, with some providers kept out of GTC.
The market figures behind the story are shifting too. According to industry data cited in the report, as of September there were more than 300 GPU providers in the neocloud space, up 55% from about 209 a year earlier, and these companies took in more than $25 billion in revenue in 2025. Nvidia's share of AI accelerator revenue is expected to fall from 87% in 2024 to around 75% this year. Neoclouds choose AMD mainly for price, while Nvidia still leads on performance.
The software layer is loosening too
In the same week, Google open-sourced TorchTPU, a PyTorch-native TPU backend, and AWS announced TorchNeuron, a native backend for Trainium, with Meta backing both efforts. What has kept developers locked into Nvidia's ecosystem is, to a large extent, CUDA and its accompanying software. When mainstream frameworks reduce a hardware switch to changing one device name, hardware choices come down more to price and availability.
That is also why neoclouds are caught between two bad options: an extra type of chip can lower costs and win more customers, but it may cost them standing with Nvidia. Whether the turned-away providers will respond publicly, and whether Nvidia will step forward to clarify, remains unanswered for now.
Sources: SemiAnalysis report, Crypto Briefing, CocoLoop, Dealroom; market share and provider counts follow the industry statistics cited in the reports, and the list of turned-away providers has not been made public.