For most of the past decade, buying AI compute meant buying Nvidia. That is no longer the only option. A wave of custom AI chips from Amazon, Google, Microsoft and Meta now runs a growing share of the world's largest AI workloads, and frontier labs are joining in. The reporting that Anthropic is exploring its own chip with Samsung is the latest sign that the accelerator market is fragmenting. This guide compares the leading custom silicon against Nvidia's Blackwell GPUs on the things that actually decide a deployment: raw performance, memory, software, cost and the lock-in you accept by choosing one.
- Custom AI chips from the hyperscalers now rival Nvidia at the chip level on FP8 throughput and memory bandwidth, though Nvidia keeps an edge on memory capacity and low-precision FP4.
- Google's TPU v7 Ironwood and AWS Trainium3 are the most credible alternatives, but each runs only inside its own cloud, so choosing one means accepting vendor lock-in.
- Nvidia's real moat is CUDA and a decade of software tooling, which most teams still build on regardless of the silicon underneath.
- Broadcom and the foundries quietly enable this whole shift, designing and fabricating the ASICs that let a software company ship a competitive accelerator.
Why hyperscalers and labs build custom AI chips
The motivation is money. Nvidia commands high margins on its data-center GPUs because demand has outstripped supply for years, and every company running large models pays that premium. A hyperscaler that designs its own accelerator can strip out the markup and tune the chip for exactly the workloads it runs. As CNBC noted in its rundown of the top AI chips, the cloud giants build custom silicon precisely to reduce their dependence on Nvidia and to control the cost of the infrastructure that increasingly drives their revenue.
There is a second driver beyond price: fit. A general-purpose GPU has to be good at everything, from training to graphics to scientific computing. A custom accelerator can drop the parts it does not need and pour that silicon budget into the operations that matter for transformer inference and training. Google has followed this logic through seven generations of its Tensor Processing Unit. Amazon built Trainium and Inferentia for training and serving. Microsoft has Maia, and Meta runs its MTIA line internally. Each part reflects the workloads of the company that made it.
The catch is that designing a leading-edge chip is brutally hard and expensive, which is why this club stayed small for so long. It takes a silicon team, a foundry relationship, advanced packaging, and years of iteration to reach a part that beats simply renting Nvidia GPUs. That barrier is the reason a hire like Anthropic's recruitment of a veteran chip engineer counts as real news rather than a footnote.
The contenders at a glance
Five families dominate the conversation in 2026. Nvidia's Blackwell generation, spanning the B200 and the higher-memory B300, remains the part everyone else is measured against and the only one you can rent across multiple clouds. Google's TPU v7, code-named Ironwood, is the most mature custom challenger and is tuned heavily for inference. AWS Trainium3 is Amazon's answer, positioned as a genuine training and inference alternative. Microsoft's Maia 200 and Meta's MTIA round out the field, though both are effectively captive to their makers.
The technical detail worth internalizing is that these parts are converging at the chip level even as they diverge on everything around the chip. On paper, the leading custom accelerators now trade blows with Blackwell on throughput and bandwidth. Where they differ sharply is in memory capacity and software support, and in whether you can actually buy access without committing to a single cloud.
It helps to separate the parts by intent. Google's TPU line has always been shaped around Google's own inference and training needs, and Ironwood leans hard toward serving models at high volume. Trainium3 is Amazon's bid to be taken seriously for large training runs as well as inference, which is why the reporting frames it against Nvidia's Blackwell Ultra rather than a lesser part. Maia and MTIA, by contrast, are not really trying to win an open market at all. They exist so that Microsoft and Meta can run their own AI services on hardware they control, which is a strategic goal rather than a product you can evaluate. Keeping that intent in mind stops the comparison from turning into a meaningless race of spec-sheet numbers between parts that were never meant to compete for the same buyer.
Chip-level specs compared
Start with raw compute. According to a detailed breakdown from Spheron, Google's TPU v7 Ironwood delivers roughly 4,614 FP8 teraflops per chip, essentially neck and neck with Nvidia's B200 at about 4,500 FP8 teraflops. AWS Trainium3 lands in the same weight class, with reporting from Tom's Hardware framing Trainium3 as competing directly against Nvidia's Blackwell Ultra in FP8 performance. At the level of a single chip, the custom parts have caught up.
Memory is where Nvidia keeps a lead. The B300 ships with 288 GB of HBM3e and about 8 TB/s of bandwidth, ahead of Ironwood's 192 GB at 7.37 TB/s and well ahead of Trainium3's 144 GB at 4.9 TB/s. Microsoft's Maia 200 sits between them with 216 GB and around 7 TB/s. More on-chip memory lets a GPU hold larger models and longer context without splitting work across chips, which is a real advantage for the biggest frontier models. Process nodes tell a similar story of parity: the custom parts are built on 3nm-class TSMC processes, while Nvidia's current Blackwell parts use a 4nm-class node, so the challengers are not behind on manufacturing.
Precision support is the subtler gap. Nvidia's Blackwell architecture supports FP4, an ultra-low-precision format that can roughly double inference throughput for models quantized to fit it. Google only added FP8 support to its TPUs in 2025, years after Nvidia, and the custom accelerators generally trail on the newest low-precision formats. For teams squeezing every token per watt out of quantized models, that head start still favors Nvidia.
Chip-level teraflop numbers are marketing-friendly but incomplete. Real performance depends on memory bandwidth, interconnect between chips, software maturity, and how well a given model maps onto the hardware. A part that wins a spec sheet can still lose on the workload you actually run, which is why rack-scale and pod-scale behavior matters as much as per-chip figures.
Rack scale changes the picture
A single chip rarely serves a frontier model on its own. What matters in a data center is how the parts scale when wired together into a rack or a pod, and here the vendors take different paths. Nvidia's GB200 NVL72 links 72 Blackwell GPUs into one tightly coupled rack that behaves like a single large accelerator, delivering on the order of 0.36 FP8 exaflops. AWS designed its Trn3 UltraServer to answer exactly this, packing 144 Trainium3 chips into a rack that reaches a comparable 0.36 exaflops of FP8 performance, roughly matching Nvidia's flagship rack.
Google plays a different game with scale. Its Ironwood pods connect thousands of TPUs through a custom optical interconnect, and a full pod of 9,216 chips reaches into the tens of exaflops. Some coverage has translated that into eye-catching multiples over a single Nvidia rack, but those comparisons put a whole pod against one rack, which is not a like-for-like matchup. The honest read is that Google can scale a single coherent TPU domain to a very large size, which suits training runs and high-volume inference, while the per-chip and per-rack gaps against Nvidia are far narrower than the pod headlines suggest.
The power wall and efficiency
There is a constraint that increasingly overshadows raw speed: electricity. New AI clusters are now planned in gigawatts of power draw rather than counts of servers, and the grid has become a binding limit on how fast any company can expand. That reality changes what a good chip even means. The metric that matters is not peak teraflops but performance per watt, because a data center with a fixed power allocation can only serve as many tokens as its efficiency allows.
This is exactly the ground custom silicon was built to win. Google positions Ironwood as an inference-optimized part whose efficiency underpins its total cost of ownership claims, and Amazon makes a parallel pitch that Trainium serves more work per dollar and per watt than renting general-purpose GPUs. Nvidia has answered with its own efficiency gains and with FP4, which lets a Blackwell GPU push more quantized inference through the same power envelope. The competition on efficiency is fierce precisely because whoever serves more intelligence per watt can either lower prices or grow faster inside the same power budget. For a lab weighing whether to design its own chip, that equation is a large part of the appeal, since a more efficient accelerator stretches the finite gigawatts it has managed to secure.
The software moat
Hardware parity does not equal parity in practice, because the software layer decides how easily a team can actually use a chip. This is Nvidia's deepest advantage. CUDA has had close to two decades of investment, and the surrounding tools for inference, such as TensorRT-LLM and broad support in open frameworks, mean most AI code runs on Nvidia with little friction. Porting is rarely needed because the ecosystem assumes Nvidia by default.
The custom chips each ask for a different toolchain. AWS Trainium runs through the Neuron SDK, Amazon's proprietary stack, and moving a workload onto it typically costs engineering time measured in weeks. Google's TPUs lean on JAX along with tools like MaxText and JetStream, which are powerful but unfamiliar to teams raised on the Nvidia and PyTorch world. That software distance is a hidden tax. A chip that is 10 percent cheaper on paper can still cost more once you count the engineers who have to rewrite and revalidate the pipeline, which is a big reason CUDA keeps its grip even as rivals close the hardware gap.
The gap has begun to narrow in the past couple of years, and that shift is the real trend to watch here. Open inference engines and portable compilers have started to smooth the path onto non-Nvidia hardware, so that serving a popular open-weight model on a TPU or a Trainium instance is less of a hand-built effort than it was two years ago. The cloud providers have strong incentives to fund this work, because every model that runs cleanly on their silicon is a customer who no longer needs to rent Nvidia capacity from a rival. If the tooling reaches the point where switching accelerators is a configuration change rather than a rewrite, the balance of power shifts toward whoever offers the best price per token. Nvidia's software lead is real and durable, but it is not the permanent moat it looked like when CUDA was the only serious game in town. The direction of the tooling is the quiet variable that could decide how fast the custom parts take share.
Cost, availability and lock-in
The strongest case for custom silicon is total cost of ownership at scale. Google has claimed that the all-in cost per Ironwood chip, across a full configuration, runs meaningfully below the cost of an equivalent Nvidia server, with figures in the range of a 40 percent-plus TCO reduction cited in vendor analysis. Amazon makes a similar efficiency pitch for Trainium. For a hyperscaler running its own models at enormous volume, savings like that compound into billions.
Availability is the flip side, and it is where the custom parts fall short for most buyers. Ironwood is available only through Google Cloud, and Trainium only through AWS. Microsoft's Maia 200 is not rentable at all: there is no public instance type and no migration path, because it exists to serve Microsoft's own services. Meta's MTIA is purely internal. Nvidia, by contrast, is available across essentially every cloud and through on-premises hardware, which is why the roughly 99 percent of teams that are not hyperscalers still default to it.
Lock-in ties these threads together. Choosing Trainium or a TPU means adopting a proprietary software stack and a single provider, with real switching costs if you later want to move. Some analysts project that Nvidia's share of inference compute could slide from around 90 percent toward the 20 to 30 percent range by the end of the decade, but that projection describes the captive workloads hyperscalers run for themselves, not the open market most companies buy in. For a typical team, Nvidia remains the safe, portable choice, and custom silicon is a bet you make only when your scale justifies the commitment.
Where Broadcom and the foundries fit
Behind every custom accelerator sits a partner that most users never see. Broadcom co-designs the silicon for several of these programs, including Google's TPUs and OpenAI's inference chip, providing the design expertise a software company lacks in-house. TSMC fabricates the overwhelming majority of leading accelerators, from Nvidia's GPUs to the custom parts, which makes its manufacturing capacity one of the true bottlenecks of the AI era. Samsung is now trying to break into that role, which is what makes its talks with Anthropic significant beyond a single chip.
This supply chain is why a custom accelerator is never just one company's achievement. A lab supplies the workload knowledge and the funding, a design partner like Broadcom translates that into a chip, and a foundry like TSMC or Samsung turns the design into physical silicon. When Anthropic weighs Samsung's advanced process for a future part, it is really assembling that same three-way relationship that Google and Amazon built years earlier. The more labs that assemble it, the more the accelerator market spreads out from a single dominant supplier.
Choosing between custom AI chips and Nvidia
The practical takeaway is that there is no single winner, only a fit for your situation. If you operate at hyperscaler volume, run your own models, and can absorb the software cost, a custom accelerator like Ironwood or Trainium3 can deliver a lower total cost of ownership than renting GPUs, which is exactly why the biggest players built them. If you are anyone else, Nvidia's combination of portable software, broad availability, and a memory and precision edge still makes it the default that is hardest to regret.
What has genuinely changed in 2026 is that the question is now live at all. A few years ago, comparing custom AI chips to Nvidia was mostly academic because the alternatives were immature or unavailable. Today the leading custom parts match Blackwell on core throughput, undercut it on cost at scale, and trail it mainly on memory, precision and the software ecosystem. That is a much closer race than the market's 90 percent Nvidia share implies, and the entry of frontier labs like Anthropic will only tighten it. The winners over the next few years will be the buyers who match the chip to the workload instead of defaulting to a single vendor out of habit.