Fierce Network reports that “many” AI neoclouds are ordering AMD and Nvidia GPUs at rough 1:1 parity, with at least two providers reportedly standing up 10,000 accelerators from each vendor side by side. If those deployments materialize at the scale described by VAST Data’s John Mao, they would represent a meaningful change in how GPU-specialist clouds hedge supplier risk. The claim, however, is an account from one executive at a company that has just expanded its AMD go-to-market partnership—not disclosed purchase orders, customer announcements, or an independently measured market-share shift.
That distinction is central. A 10,000-GPU AMD cluster paired with 10,000 Nvidia GPUs says a provider wants a credible second supply line. It does not demonstrate equal AI capacity, equal customer demand, equal revenue, or a broad market split. The GPU models, memory configurations, networking, power density, delivery dates and software stacks were not identified, and neither VAST nor AMD has named the two neoclouds Mao referenced.
VAST has a direct commercial interest in proving that its AI Operating System can sit beneath either vendor’s accelerators. Its July 23 announcement with AMD presents the partnership as a way for AI clouds to build training and inference systems around AMD EPYC CPUs, Instinct GPUs, Pensando networking, ROCm and VAST’s data layer. That makes Mao’s remarks important market color, but it also means they should be read as evidence of buyer exploration and infrastructure diversification, rather than a settled verdict on Nvidia’s position.
Fierce Network’s reporting is the only identified source for the 1:1 ordering claim, and it relies on Mao’s conversations with providers at AMD Advancing AI 2026. Mao said one neocloud buyer viewed Nvidia as holding the performance edge while AMD offered a price advantage. That is a plausible procurement argument, especially when GPU availability is constrained, but VAST did not disclose the price difference, the workload mix, or whether the clusters are intended for training, inference, reserved-capacity contracts, internal use, or retail GPU cloud instances.
The missing product details matter. Counting accelerators is an unusually weak way to compare competing AI fleets when the comparison could span different Nvidia and AMD generations. A fleet of Nvidia GPUs configured for high-end training may have very different throughput, memory behavior, rack power requirements and customer economics from an AMD fleet aimed at inference-heavy workloads. Even within the same vendor, accelerator count does not reveal whether the machines are eight-GPU systems, rack-scale designs, leased capacity, or hardware awaiting deployment.
There is also a difference between standing up 10,000 GPUs and having them fully operational, rented, and generating repeatable revenue. Neoclouds have aggressively reserved chips, power and data-center space in advance of customer demand because late capacity can mean losing a large AI contract. A procurement decision therefore signals confidence and supply strategy more clearly than it signals which platform enterprise customers will choose.
The immediate read-through for IT buyers is narrower but practical: AMD is increasingly likely to be available as a real option at GPU-cloud providers that previously offered Nvidia almost exclusively. That gives customers more room to negotiate availability, contract terms and software support. It does not remove the need to benchmark the specific model, framework and inference stack being purchased.
But Fierce Network’s summary contains an important reporting error: AMD did not project that Instinct sales would rise 46% in the second quarter. The 46% figure was AMD’s forecast for total company revenue growth at the midpoint of its Q2 outlook—approximately $11.2 billion in revenue—not an Instinct-specific growth forecast.
AMD’s earnings-call record separates the two points. It gives the more-than-70% outlook for server CPU revenue, while describing data-center AI as expected to grow sequentially by double digits along with the server business. AMD did not publish a standalone Q2 Instinct revenue estimate. As of August 3, AMD has also not reported its second-quarter results; it is scheduled to do so after the market closes on August 4.
That correction does not diminish AMD’s momentum. It does change the certainty attached to the story. AMD’s $5.8 billion Data Center figure combines CPU and GPU business, so it cannot on its own establish how much of the gain came from Instinct accelerators or whether neocloud orders are the main driver. The company has announced major future AI infrastructure commitments, including Meta’s plan to deploy up to 6 gigawatts of AMD Instinct GPUs over several product generations, but those announcements are separate from the unnamed neocloud deployments described by Mao.
For administrators and infrastructure planners, the lesson is to distinguish vendor revenue growth from platform maturity in a specific environment. AMD’s data-center growth confirms more customers are buying its hardware. It does not prove that an existing CUDA estate can move workloads to ROCm without engineering cost, retraining, framework validation and operational changes.
That is relevant to AI infrastructure because a GPU cluster rarely fails on floating-point performance alone. Model loading, checkpointing, retrieval-augmented generation, event streams, persistent context and KV cache handling can leave expensive accelerators waiting for data. VAST is trying to make its storage and data services a common layer that can serve AMD-based and Nvidia-based clusters, reducing the penalty for a neocloud that carries both.
The company’s own benchmark claims should still be treated carefully. VAST says early MI355X testing showed a ninefold time-to-first-token improvement and 9.7 times higher token throughput with KV-cache offload in high-concurrency agentic workloads. Its release also says the performance and cost claims were reviewed but not independently verified by AMD, and that results may not be typical.
That disclosure is unusually useful because it identifies what remains unproven. The benchmark may demonstrate that VAST’s architecture improves a chosen baseline under a chosen workload. It does not establish a general MI355X advantage over Nvidia hardware, nor does it show what an organization will see with a different model, context length, storage tier, network fabric or concurrency profile.
The partnership’s real value is therefore operational rather than promotional: it gives AMD-based cloud builders an integrated storage, networking and reference-architecture story at a time when customers are demanding more than bare GPU rental. It also gives VAST a reason to promote a multi-vendor accelerator market. Those incentives align, but they are not the same as independent confirmation that AMD and Nvidia have reached parity in neocloud purchasing.
The forecast should not be mistaken for guaranteed demand, but the current direction helps explain why providers would build heterogeneous fleets. A neocloud’s business is constrained by power, data-center capacity, financing, networking and chips. When supply of any one component is scarce, a provider that can sell both Nvidia- and AMD-backed capacity gains more ways to turn an energized rack into revenue.
Nvidia still benefits from the installed base, CUDA familiarity and mature software expectations that make it the default platform for many teams. Mao himself acknowledged Nvidia’s performance advantage in the provider conversation he cited. AMD’s opportunity is not necessarily to displace Nvidia in a single purchasing cycle; it is to become sufficiently deployable, supportable and available that buyers no longer accept a single-vendor fleet as inevitable.
That transition will be visible through public customer disclosures, production cloud-instance availability, framework support and utilization—not through an unnamed 1:1 order ratio alone. AMD’s August 4 earnings report is the next near-term checkpoint for revenue evidence. The larger proof will come when the unnamed 10,000-GPU deployments are identified, the GPU configurations are known, and customers show that they are consuming AMD capacity at rates comparable to the Nvidia systems beside it.
VAST has a direct commercial interest in proving that its AI Operating System can sit beneath either vendor’s accelerators. Its July 23 announcement with AMD presents the partnership as a way for AI clouds to build training and inference systems around AMD EPYC CPUs, Instinct GPUs, Pensando networking, ROCm and VAST’s data layer. That makes Mao’s remarks important market color, but it also means they should be read as evidence of buyer exploration and infrastructure diversification, rather than a settled verdict on Nvidia’s position.
The 1:1 figure is an order claim, not a market-share measurement
Fierce Network’s reporting is the only identified source for the 1:1 ordering claim, and it relies on Mao’s conversations with providers at AMD Advancing AI 2026. Mao said one neocloud buyer viewed Nvidia as holding the performance edge while AMD offered a price advantage. That is a plausible procurement argument, especially when GPU availability is constrained, but VAST did not disclose the price difference, the workload mix, or whether the clusters are intended for training, inference, reserved-capacity contracts, internal use, or retail GPU cloud instances.The missing product details matter. Counting accelerators is an unusually weak way to compare competing AI fleets when the comparison could span different Nvidia and AMD generations. A fleet of Nvidia GPUs configured for high-end training may have very different throughput, memory behavior, rack power requirements and customer economics from an AMD fleet aimed at inference-heavy workloads. Even within the same vendor, accelerator count does not reveal whether the machines are eight-GPU systems, rack-scale designs, leased capacity, or hardware awaiting deployment.
There is also a difference between standing up 10,000 GPUs and having them fully operational, rented, and generating repeatable revenue. Neoclouds have aggressively reserved chips, power and data-center space in advance of customer demand because late capacity can mean losing a large AI contract. A procurement decision therefore signals confidence and supply strategy more clearly than it signals which platform enterprise customers will choose.
The immediate read-through for IT buyers is narrower but practical: AMD is increasingly likely to be available as a real option at GPU-cloud providers that previously offered Nvidia almost exclusively. That gives customers more room to negotiate availability, contract terms and software support. It does not remove the need to benchmark the specific model, framework and inference stack being purchased.
AMD’s financials support demand, but not the claimed Instinct forecast
AMD’s first-quarter 2026 results provide hard evidence that its data-center business is expanding. The company reported $5.8 billion in Data Center revenue, up 57% from the prior year, crediting EPYC server processors and the continuing ramp of Instinct GPU shipments. AMD also said it expected server CPU revenue to grow more than 70% year over year in the second quarter.But Fierce Network’s summary contains an important reporting error: AMD did not project that Instinct sales would rise 46% in the second quarter. The 46% figure was AMD’s forecast for total company revenue growth at the midpoint of its Q2 outlook—approximately $11.2 billion in revenue—not an Instinct-specific growth forecast.
AMD’s earnings-call record separates the two points. It gives the more-than-70% outlook for server CPU revenue, while describing data-center AI as expected to grow sequentially by double digits along with the server business. AMD did not publish a standalone Q2 Instinct revenue estimate. As of August 3, AMD has also not reported its second-quarter results; it is scheduled to do so after the market closes on August 4.
That correction does not diminish AMD’s momentum. It does change the certainty attached to the story. AMD’s $5.8 billion Data Center figure combines CPU and GPU business, so it cannot on its own establish how much of the gain came from Instinct accelerators or whether neocloud orders are the main driver. The company has announced major future AI infrastructure commitments, including Meta’s plan to deploy up to 6 gigawatts of AMD Instinct GPUs over several product generations, but those announcements are separate from the unnamed neocloud deployments described by Mao.
For administrators and infrastructure planners, the lesson is to distinguish vendor revenue growth from platform maturity in a specific environment. AMD’s data-center growth confirms more customers are buying its hardware. It does not prove that an existing CUDA estate can move workloads to ROCm without engineering cost, retraining, framework validation and operational changes.
VAST’s AMD partnership targets the data path around the GPU
The more concrete announcement is VAST and AMD’s July partnership. VAST selected sixth-generation AMD EPYC processors, formerly codenamed Venice, for the sixth generation of its CBox platform and the third generation of EBox. The stated hardware benefit is PCIe Gen 6 support, which VAST says doubles generational I/O bandwidth and reduces latency for its database, data warehouse and event-streaming services.That is relevant to AI infrastructure because a GPU cluster rarely fails on floating-point performance alone. Model loading, checkpointing, retrieval-augmented generation, event streams, persistent context and KV cache handling can leave expensive accelerators waiting for data. VAST is trying to make its storage and data services a common layer that can serve AMD-based and Nvidia-based clusters, reducing the penalty for a neocloud that carries both.
The company’s own benchmark claims should still be treated carefully. VAST says early MI355X testing showed a ninefold time-to-first-token improvement and 9.7 times higher token throughput with KV-cache offload in high-concurrency agentic workloads. Its release also says the performance and cost claims were reviewed but not independently verified by AMD, and that results may not be typical.
That disclosure is unusually useful because it identifies what remains unproven. The benchmark may demonstrate that VAST’s architecture improves a chosen baseline under a chosen workload. It does not establish a general MI355X advantage over Nvidia hardware, nor does it show what an organization will see with a different model, context length, storage tier, network fabric or concurrency profile.
The partnership’s real value is therefore operational rather than promotional: it gives AMD-based cloud builders an integrated storage, networking and reference-architecture story at a time when customers are demanding more than bare GPU rental. It also gives VAST a reason to promote a multi-vendor accelerator market. Those incentives align, but they are not the same as independent confirmation that AMD and Nvidia have reached parity in neocloud purchasing.
Supply constraints may be forcing dual-sourcing faster than software preference would
Mao told Fierce Network that lead-time pressure has worsened rather than eased over the last 12 months for Nvidia, AMD and VAST alike. That observation is consistent with the broader neocloud expansion documented by Synergy Research Group: the analyst firm put neocloud revenue above $25 billion in 2025 and forecast a market approaching $400 billion by 2031.The forecast should not be mistaken for guaranteed demand, but the current direction helps explain why providers would build heterogeneous fleets. A neocloud’s business is constrained by power, data-center capacity, financing, networking and chips. When supply of any one component is scarce, a provider that can sell both Nvidia- and AMD-backed capacity gains more ways to turn an energized rack into revenue.
Nvidia still benefits from the installed base, CUDA familiarity and mature software expectations that make it the default platform for many teams. Mao himself acknowledged Nvidia’s performance advantage in the provider conversation he cited. AMD’s opportunity is not necessarily to displace Nvidia in a single purchasing cycle; it is to become sufficiently deployable, supportable and available that buyers no longer accept a single-vendor fleet as inevitable.
That transition will be visible through public customer disclosures, production cloud-instance availability, framework support and utilization—not through an unnamed 1:1 order ratio alone. AMD’s August 4 earnings report is the next near-term checkpoint for revenue evidence. The larger proof will come when the unnamed 10,000-GPU deployments are identified, the GPU configurations are known, and customers show that they are consuming AMD capacity at rates comparable to the Nvidia systems beside it.
References
- Primary source: Fierce Network
Published: 2026-08-03T14:00:00+00:00
AMD is 1:1 with Nvidia in many neoclouds, says VAST exec
Many neoclouds are deploying AMD Instinct GPUs at parity with Nvidia, says VAST Data's John Mao. And VAST expanded its AMD partnership to serve both fleets.www.fierce-network.com - Related coverage: telecompaper.com
Neoclouds attract 223% more turnover in Q4 - Synergy - Telecompaper
Revenues from neoclouds, meaning GPU-specialised cloud products and services, reached USD 9 billion in the final quarter of 2025, an increase of 223 percent year-on-year, according to Synergy Research Group. Full year sales surpassed USD 25 billion globally, driven by surging demand for AI...www.telecompaper.com - Related coverage: datacentermarket.es
Los neoclouds de IA valdrán 400.000 millones de dólares en 2031 | Data Center Market
Impulsados por una demanda de GPU que los gigantes tradicionales no logran satisfacer, los proveedores neocloud han pasado de ser un nichowww.datacentermarket.es - Related coverage: vastdata.com
VAST Data Valued at $30 Billion as AI Drives a New Infrastructure Stack
VAST delivers the first AI Operating System, unifying storage, database, and compute to drive agentic computing and data intensive workloads. Learn more.www.vastdata.com - Related coverage: storagereview.com
VAST Data and AMD Claim 9x Faster Time-to-First-Token With KV Cache Offload on Instinct - StorageReview.com
VAST Data expands its AMD collaboration: 6th Gen EPYC CBox/EBox, a Helios reference architecture, and 9x faster TTFT via KV cache offload.
www.storagereview.com
- Related coverage: news.bloomberglaw.com
Nvidia-Backed Vast Data Triples Value to $30 Billion, Mulls IPO
Nvidia Corp.-backed Vast Data Inc., which makes software for storing data for artificial intelligence tasks, said it has raised about $1 billion and more than tripled its valuation to $30 billion.news.bloomberglaw.com
- Related coverage: hpcwire.com
HPCwire - Since 1987 – Covering the Fastest Computers in the World and the People Who Run Them
July 24, 2026 – VAST Data has announced an expanded collaboration with AMD to help AI cloud providers and enterprises build and operate high-performance AI factories at scale. The collaboration brings together the VAST AI Operating System with 6th Gen AMD EPYC CPUs and AMD Instinct GPUs to...www.hpcwire.com - Related coverage: bloomberg.com
