Custom silicon from AWS, Google, and Microsoft has moved from R and D curiosity to production scale infrastructure, running real training and inference workloads in 2025 and 2026.
For most of the current AI buildout, renting NVIDIA GPUs has been the default and, for many teams, the only practical option. That default is cracking. AWS, Google Cloud, and Microsoft Azure have each spent years and billions of dollars building their own AI accelerators, and by 2026 those chips are no longer side projects. They are running production training and inference workloads for some of the largest AI companies in the world, a shift also visible in how GPU supply and demand has been fragmenting across the market.
The motivation is straightforward economics. NVIDIA's gross margins on data center GPUs have consistently run above 70 percent, according to the company's own quarterly earnings disclosures, and that margin is effectively a tax every renter pays on top of the underlying silicon cost. A hyperscaler that designs its own chip, even at a higher unit engineering cost, can capture that margin internally and pass some of the savings to anchor customers willing to commit to the ecosystem.
This matters for infrastructure buyers beyond the hyperscalers themselves. Custom silicon changes the negotiating position of every enterprise buying GPU capacity, because it introduces a credible alternative that did not meaningfully exist five years ago. Even teams that never touch a Trainium or TPU chip benefit from the price pressure this competition puts on NVIDIA based instance pricing.
AWS's Trainium line is the clearest example of a hyperscaler treating custom silicon as core infrastructure rather than an experiment. Trainium2 became generally available in late 2024, and AWS has stated the chip delivers roughly four times the performance of its first generation Trainium instances, with up to 30 to 40 percent better price performance than current GPU based EC2 instances, according to AWS's own re:Invent announcements.
The most significant proof point is Project Rainier, the Trainium2 based cluster AWS built for Anthropic. Reporting from Reuters and CNBC in 2025 described the cluster as spanning hundreds of thousands of Trainium2 chips, making it one of the largest dedicated AI training deployments outside of NVIDIA based infrastructure. Anthropic's willingness to train frontier models on non NVIDIA silicon is the strongest signal yet that custom accelerators can handle workloads previously assumed to require merchant GPUs.
Trainium3, previewed at re:Invent 2024 for a 2025 to 2026 rollout, is built on a more advanced process node and is positioned by AWS as the chip that closes the remaining software and performance gap with NVIDIA's Blackwell generation. The open question for buyers is not raw performance. It is whether the software stack, primarily AWS's Neuron SDK, has matured enough that migrating a training pipeline away from CUDA is a manageable engineering project rather than a multi quarter rewrite.
Google has the longest track record here. Tensor Processing Units have been in production since 2015, originally built for internal workloads like Search and Translate before Google Cloud opened them to external customers. That decade of iteration shows in the software maturity: the JAX and XLA toolchains built around TPUs are further along than any competing custom silicon stack, precisely because Google has been dogfooding its own hardware internally for longer than any rival.
Trillium, Google's sixth generation TPU, reached general availability in late 2024 with a claimed 4.7 times improvement in peak compute performance per chip over the fifth generation TPU v5e, according to Google Cloud's own product announcements. At Google Cloud Next 2025, Google introduced Ironwood, its seventh generation TPU, explicitly positioned for large scale inference rather than training, reflecting the industry wide shift in where AI compute spend is actually going as more workloads move from model development into production serving.
TPUs remain a Google Cloud exclusive; there is no path to renting TPU capacity through any other provider, and no path to buying the chips outright. That is both the constraint and the pitch. Buyers get the most mature non NVIDIA software stack available today, in exchange for full commitment to Google Cloud's ecosystem for that workload.
Microsoft's Maia 100 accelerator, announced at Ignite in November 2023, has taken a more conservative rollout path than Trainium or TPU. Microsoft has used Maia primarily for internal workloads, including portions of Copilot and Azure OpenAI Service infrastructure, rather than aggressively pushing external customers onto it. That caution reflects a genuine tension in Microsoft's position: Azure's single largest AI customer relationship, with OpenAI, runs overwhelmingly on NVIDIA hardware today, and Microsoft has less incentive than AWS or Google to force a rapid migration away from a stack that is already working at scale.
The strategic logic for Maia is less about immediate cost savings on any single customer and more about supply diversification. Every accelerator Microsoft can source internally is one less GPU it needs to secure from NVIDIA's constrained allocation, which matters when demand for AI compute continues to outstrip supply industry wide. For buyers, this means Maia is currently the least externally accessible of the three major custom silicon programs and the one to watch rather than the one to plan around in the near term.
The hyperscalers are not the only companies designing their own chips. Reuters reported in 2025 that OpenAI is working with Broadcom to design a custom AI accelerator for internal use, targeting deployment in 2026, following a pattern already established by Meta's MTIA, or Meta Training and Inference Accelerator, chips, which handle a portion of Meta's recommendation and ranking workloads today.
This is a meaningful signal for anyone tracking where GPU demand is headed. If the largest AI labs and hyperscalers are all moving portions of their workloads to chips they control end to end, that reduces the addressable market NVIDIA competes for at the very top of the demand curve, even as overall AI compute demand keeps climbing. It also reinforces a theme covered elsewhere on this site: the GPU supply picture is shifting from pure scarcity toward a more fragmented, multi architecture landscape, and buyers with negotiating leverage should expect that fragmentation to show up in pricing.
The headline price performance claims from AWS and Google are vendor reported and should be treated as directional rather than verified, consistent with how this site treats any vendor sourced benchmark. The more useful framing for a buyer is where the savings actually come from, and where they do not.
Savings tend to show up in raw compute cost per token or per training step, because the hyperscaler is not paying NVIDIA's margin on the underlying silicon. Savings do not reliably show up in total delivered cost, because migrating a workload to Trainium or TPU carries real engineering cost: re validating a training pipeline, retuning distributed training code for a different interconnect topology, and in some cases rewriting custom CUDA kernels that have no direct equivalent on the new hardware. Teams that underestimate this migration cost tend to end up with a bill that looks similar to what they were paying before, just redistributed from compute spend to engineering spend. Running the numbers through a cloud cost calculator before committing to a migration is the fastest way to see whether the compute savings survive contact with the engineering bill.
The workloads best suited to a first custom silicon experiment are large, well understood, and running on frameworks with strong native support: standard transformer training on JAX for TPU, or well established PyTorch training loops for Trainium using AWS's Neuron SDK. Workloads that lean heavily on custom CUDA kernels, exotic model architectures, or third party libraries with no non NVIDIA support are the wrong place to start.
| Program | Owner | Primary Use Today | External Availability |
|---|---|---|---|
| Trainium2 / 3 | AWS | Training and inference, proven at Project Rainier scale | AWS EC2 Trn instances |
| TPU (Trillium, Ironwood) | Training and large scale inference | Google Cloud only | |
| Maia 100 | Microsoft | Primarily internal (Copilot, Azure OpenAI infra) | Limited external access |
| MTIA | Meta | Internal recommendation and ranking workloads | Not externally available |
Before committing engineering time to a custom silicon migration, run the workload through four questions.
Training workloads benefit most from TPU's mature multi chip pod interconnect and from Trainium2's proven large cluster performance in deployments like Project Rainier. Inference workloads, where latency and cost per request dominate, are increasingly the target for newer chips like Ironwood, and the calculus is different because inference serving is more sensitive to software maturity and less tolerant of migration downtime. This is the same training versus inference distinction covered in this site's guide to choosing compute for AI workloads.
A team on vanilla PyTorch or JAX with standard model architectures has a straightforward migration path. A team with hand tuned CUDA kernels, custom fused operators, or dependencies on NVIDIA specific libraries like cuDNN extensions should expect the migration to be materially harder than any vendor's marketing suggests.
Custom silicon economics improve with scale. A team renting a handful of instances for experimentation is unlikely to see meaningful savings after accounting for engineering time. A team planning a sustained, large scale training run, the kind of commitment Anthropic made with Project Rainier, is in the range where custom silicon savings can be substantial.
Moving compute to a different hyperscaler's custom chip often means moving data too, and egress fees on that data transfer can quietly erase savings that looked compelling on paper. This is the same architecture problem covered in this site's reporting on storage as the anchor of the AI infrastructure stack: any GPU or accelerator migration should be planned alongside a storage strategy that minimizes the cost of moving data between providers, not after the compute decision is already locked in.
Custom silicon is no longer a hyperscaler science project. It is a live procurement option for any team with workloads large enough to justify the migration engineering, and its mere existence is reshaping the price NVIDIA based capacity commands across the market. Teams that ignore it entirely are leaving negotiating leverage on the table. Teams that chase it without an honest accounting of migration cost, software maturity, and data movement will find the savings are smaller than the slide deck promised.
The question for 2026 is no longer whether alternatives to NVIDIA exist. It is whether your workload, your software stack, and your data are actually portable enough to use them.
Compare AWS, Google Cloud, Azure, and alternatives like Backblaze B2 Discover how much you could save in seconds
