- A July 2026 VentureBeat survey of 170 AI infrastructure respondents found 39.4% plan to evaluate non-Nvidia accelerators in the next 12 months, 14 percentage points more than those considering Nvidia’s next-gen Blackwell GPUs.
- HBM and CoWoS packaging bottlenecks are expected to persist into mid-2027with enterprise B200 orders placed in May 2026 facing 12 to 18-month lead times, pushing procurement teams toward AMD Instinct, custom ASICs and neocloud providers.
- A June 2026 Broadcom survey found 56% of enterprises are running or planning production inferencing in private cloud, a 15-percentage-point decline for public cloud as the primary environment year-over-year.
Nvidia still dominates production AI infrastructure, but a July 2026 VentureBeat survey suggests enterprises are actively building alternatives into their plans. Of 170 AI infrastructure respondents, 39.4% said they were likely to evaluate non-Nvidia accelerators within 12 months, compared to just 25.3% planning to evaluate Nvidia’s own next-generation Blackwell line. Supply constraints, cost pressure and maturing alternatives are all driving that gap.
Why Supply Stays Tight
The shortage of high-performance GPUs is structural, not cyclical. Two chokepoints dominate: High Bandwidth Memory (HBM) supply and TSMC’s CoWoS advanced packaging capacity. SK Hynix, a primary HBM supplier for Nvidia’s data centre products, has not kept pace with demand, while TSMC’s CoWoS lines remain fully allocated through at least mid-2027 despite successive capacity expansions in 2025 and 2026. By April 2026, lead times for H100 SXM5 nodes from resellers had stretched to 36 to 52 weeks.
Hyperscalers made the situation harder for everyone else. Microsoft Google Meta and Amazon placed multi-billion-dollar forward orders for Blackwell GPUs, the GB200 and B200, in 2025, locking up most of Nvidia’s available allocation through 2026 and into 2027. Mid-market and enterprise buyers placing volume B200 orders as of May 2026 were looking at 12 to 18-month lead times. That dynamic is what makes the VentureBeat diversification numbers credible: enterprises cannot simply buy their way onto the Nvidia roadmap on a normal procurement cycle.
Custom Silicon at Scale
Google (TPU v7 Ironwood), Microsoft (Maia 200), Amazon (Trainium 3) and Meta (MTIA) each run multi-generation custom accelerator programmes aimed primarily at inference workloads, which now consume roughly two-thirds of all AI compute. OpenAI has entered a 10 GW custom accelerator partnership with Broadcom, with deployment scheduled for the second half of 2026. According to Bloomberg Intelligence data from February 2026, the custom AI accelerator market is projected to grow at a 44.6% CAGR through 2033, against 16.1% for GPU-based solutions over the same period.
The logic for custom ASICs is straightforward: a chip designed for transformer inference at scale can deliver better performance-per-watt and memory bandwidth than a general-purpose GPU at the same node, without competing on the same constrained supply chain. The edge deployment market is seeing the same pattern where purpose-built silicon is displacing general-purpose GPUs for latency-sensitive workloads.
AMD and Intel Fill the Gap
Oracle plans to deploy 50,000 AMD Instinct GPUs in 2026, with further expansion planned for 2027. Meta, OpenAI and Anthropic have also announced multi-gigawatt AMD Instinct deployments starting this year. The MI300X rents at roughly 30 to 40% below H100 pricing for comparable inference performance, according to AMD, making it the most commercially visible alternative for cost-sensitive procurement teams. AMD’s MI350P PCIe cards, introduced in May 2026, target enterprises that need inference and RAG pipeline capacity without a full data centre overhaul, fitting within standard air-cooled server environments.
Microsoft is deploying the AMD Helios Rackscale Solution, combining AMD Instinct MI455X GPUs with AMD EPYC “Venice” CPUs, in Azure for frontier model inference workloads in the second half of 2026. Intel’s Gaudi 3 accelerators are available through Dell AI Factory and on IBM Cloud as of March 2025, offering another on-ramp for enterprises looking to diversify without committing to hyperscale custom silicon. The competitive picture across these alternatives is covered in detail in our Intel Gaudi 3, AMD MI300X and TPU comparison.
Cloud, Neocloud and Repatriation
Specialised GPU cloud providers, neoclouds, have carved out a pricing advantage that is now large enough to shape procurement decisions. In August 2026, median H100 rates on these platforms averaged around $4.17 per GPU hour, compared to roughly $7.89 per GPU hour on hyperscalers for the same silicon. CoreWeave, Lambda Labs and packet.ai are among the providers in this segment. CoreWeave secured a $2 billion investment from Nvidia and a $6 billion cloud agreement with Jane Street in January 2026, a signal that the segment has moved past the experimental stage.
On-premise H100 deployments are also back in the cost conversation. For sustained workloads with moderate-to-high utilisation, owned hardware amortises against cloud rental within 6 to 14 months where the enterprise has capital, data centre space, adequate power and cooling, and the engineering capacity to run the cluster. A June 2026 Broadcom survey of global enterprises found 56% are running or planning production inferencing in private cloud, a 15-percentage-point shift away from public cloud as the primary environment compared to the prior year. Security, cost control and governance are the cited drivers. Whether that repatriation holds as hyperscaler pricing responds remains an open question, but the procurement calculus has changed enough that it is no longer a default in either direction.
