0

Everyone’s making custom ASICs — so why all the talk about Nvidia dependency?

Nvidia arguably powered the AI boom, but well and truly into it, hyperscalers and other big players aren’t exactly thrilled with their reliance on the company. The major hyperscalers and frontier AI labs continue to lean on Nvidia for frontier model training and a large share of high-end inference, even as they spend billions trying to change that.

While Nvidia still functions as a de facto infrastructure monopoly for general-purpose AI compute, the big players no longer treat it as the single foundation of their future AI stacks. Microsoft, Google, Amazon, Meta, plus labs like OpenAI and Anthropic, are increasingly building around a multi-vendor, multi-architecture strategy in which Nvidia is one supplier among several. A very large supplier, to be clear — but one supplier nonetheless.

The dependence is still very much there, but it’s narrowing quickly. Custom AI chips and alternative accelerators are ramping fast enough that Wall Street has started flagging hyperscaler silicon as a genuine risk to Nvidia’s long-term dominance. The question isn’t whether hyperscalers want out from under Nvidia — they clearly do. It’s how fast they can actually get there, and how much of the stack they can realistically replace.

Scale of hyperscaler AI spending and Nvidia exposure

Reports for 2026 put aggregate AI-related capex across the major providers over $700 billion. Alphabet alone is reported in the $175–185 billion band for the year. In 2025, Amazon, Microsoft and Google collectively spent around $305 billionin capex — and a large chunk of that flowed straight to Nvidia GPU systems.

Nvidia has also locked in the future. It has reportedly booked about $95 billion in long-term supply and partnership agreements with OpenAI, Anthropic, CoreWeave and Meta, tying multi-year GPU pipelines to hyperscaler-scale deployments. Set against this, Google’s TPUs and AWS’s Trainium are real and growing — but Nvidia still commands the vast majority of data-center spend on computing systems compared to internal silicon. 

Why hyperscalers want to reduce Nvidia dependence

The motivations aren’t mysterious. Start with pricing power. Nvidia’s effective monopoly over high-end accelerators, systems and software has given it strong margin leverage, and hyperscalers see obvious strategic risk in relying on a single vendor for a foundational input when their own businesses depend on scaling AI across search, ads, cloud, consumer apps and enterprise services. Nobody wants their entire growth story bottlenecked by one supplier’s allocation decisions.

Supply constraints sharpen that worry. GPU and high-bandwidth memory shortages have repeatedly left hyperscalers with models ready to ship but no chips to run them at scale. HBM supply in particular is concentrated, with SK Hynix controlling 58% of the market and Samsung and Micron at around 21% each, as of Q1 2026, according to Counterpoint. That bottleneck limits Nvidia’s ability to scale deliveries even when demand is strong, which means tethering yourself to Nvidia means tethering yourself to a memory supply chain you don’t control.

Then there’s the economics. At hyperscaler scale custom ASICs tuned to stable model architectures can deliver up to around 65% better total cost of ownership than general-purpose GPUs for inference, according to some estimates. The upfront non-recurring engineering cost of designing a custom chip generally pays back within 18–24 months, after which the savings in power, cooling and hardware capex compound. Custom silicon also lets these companies optimize for their own proprietary models and tighten vertical integration from hardware up through compilers and frameworks, which becomes a moat in itself. 

Custom AI ASIC shipments from cloud providers are forecast to grow about 44–45% in 2026, against roughly 16% growth for merchant GPUs, with ASIC shipments projected to triple by 2027 versus 2024 levels. 

Nvidia’s enduring strengths and ecosystem lock-in

None of this means Nvidia is in trouble. Its biggest moat isn’t the silicon — it’s the software. CUDA, along with cuDNN, NCCL and TensorRT, remains deeply embedded in the AI research and engineering community. Frontier model frameworks like variants of PyTorch, JAX and TensorFlow are heavily optimized for Nvidia hardware, which makes Nvidia the path of least resistance for most developers. Hyperscalers can afford to build bespoke compiler stacks for their own chips. The external ecosystem of model developers, enterprises and startups mostly just expects Nvidia-compatible infrastructure, and that expectation is sticky.

Nvidia also doesn’t only sell chips. For a hyperscaler that needs to spin up a massive cluster quickly, buying a complete Nvidia system is often faster and lower-risk than co-designing with another vendor. That convenience is worth a premium.

Pace of innovation is the other piece. Nvidia is iterating fast, with Blackwell followed quickly by Rubin, pushing performance per watt and per dollar for training and increasingly for inference. Compared to AMD’s MI300 series or Intel’s Gaudi accelerators, Nvidia’s architectural cadence is difficult to keep up with, and it keeps the company ahead on training in particular. So long as Nvidia holds the performance lead for general-purpose training, hyperscalers will keep anchoring frontier workloads on its clusters even while they offload inference elsewhere. And Nvidia’s demand is broadening — beyond hyperscalers to enterprises, sovereign clouds, telcos and sector-specific AI factories. That diversification makes it harder for the big clouds to squeeze Nvidia on price, because someone else is picking up the slack.

Key hyperscaler silicon strategies

Each of the Big Four is pursuing the same broad idea in its own way. Google has the longest head start, having invested in TPUs for nearly a decade and now treating them as a first-class alternative to Nvidia internally. Newer generations target both training and inference, often tuned for Gemini, search ranking and YouTube recommendations, and Google Cloud offers them to customers as a cost-effective option for certain workloads. Google sells its TPUs to others too. Still, Google buys significant Nvidia volume, particularly for frontier training and third-party workloads that need CUDA compatibility.

AWS has built out the Inferentia and Trainium families, with its most recent models aimed at beating Nvidia on price-performance for high-volume inference like Alexa, ads and retail recommendations. Amazon uses these internally and offers them via EC2, while still buying large quantities of Nvidia GPUs for frontier training and for customers who insist on Nvidia tooling. 

Microsoft’s approach is multi-pronged, pairing the custom Arm-based Cobalt CPU with the Maia AI accelerator co-designed with external silicon partners. But Microsoft remains, by many analysts’ reckoning, Nvidia’s single largest hyperscaler customer, thanks to its role as OpenAI’s primary cloud partner and the sheer scale of Azure’s AI services across Copilot, Bing, GitHub and Office. The OpenAI relationship is itself a major Nvidia demand driver — the two have partnered on plans for at least 10 gigawatts of AI data-center capacity built on Nvidia technology. Maia is meant to improve internal efficiency, provide negotiating leverage, and offer differentiated Azure SKUs, but Copilot, OpenAI training and most enterprise workloads stay anchored on Nvidia clusters for now.

Meta has gone hard on the MTIA accelerator to optimize recommendation, ranking and generative AI across Facebook, Instagram, WhatsApp and Threads, and has signaled 30–40% better price-performance for internal workloads versus general GPUs. It’s also been a massive Nvidia buyer for its LLaMA models. Meta’s path, like the others, is dual-track — Nvidia for frontier training and ecosystem-compatible external workloads, MTIA for high-volume internal inference. Across all four, custom chips are increasingly dominant for internal inference and emerging as meaningful training contributors, which is exactly why analysts now flag hyperscaler silicon as a significant risk to Nvidia’s pricing power.

Major AI labs: OpenAI and Anthropic

The labs aren’t traditional hyperscalers, but their compute footprints are hyperscaler-scale and central to Nvidia’s business, so they belong in this conversation. OpenAI sits mostly on Microsoft Azure, and its partnership with Nvidia is deep — those 10-plus gigawatts of planned capacity, plus preferential access to new architectures for frontier training. At the same time, OpenAI has built a custom AI chip called Jalapeño, tailored explicitly to its own models, which is already running for inference workloads. The likely future is early frontier training still leaning on Nvidia for ecosystem and flexibility, with steady-state inference shifting partly to OpenAI-specific ASICs integrated tightly with Azure.

Anthropic’s story is more cloud-mediated. It has major strategic investments and cloud deals with both AWS and Google Cloud, and its current deployments lean heavily on Nvidia GPUs across those clouds for training Claude. Anthropic is also part of Nvidia’s long-term supply commitments. But it’s emerging as an early adopter of AWS Trainium and Google TPUs for inference and some training, pushed there by cloud incentives and cost pressure. Of the two labs, Anthropic looks further along in actually diversifying away from Nvidia hardware, largely because its cloud partners are actively steering it toward their own chips. Both labs still train frontier models on Nvidia, though.

Alternative merchant chip suppliers

Beyond custom silicon, there’s a merchant ecosystem trying to be the second source. AMD’s MI300 series and its successors have begun winning share at certain hyperscalers and neoclouds, with at least one — TensorWave — reportedly building its entire AI network on AMD chips. Hyperscalers view AMD as a credible second source and, just as usefully, a bargaining chip to extract better terms from Nvidia. AMD still lacks Nvidia’s ecosystem depth and software maturity, but it’s closing performance gaps on specific workloads and cost-sensitive deployments. For now it functions more as a lever than a replacement.

Intel’s Gaudi accelerators target cloud and enterprise AI with cost-competitive configurations, though adoption at hyperscaler scale remains more limited than either Nvidia or AMD. Gaudi is another second-source option rather than a serious challenger. The more interesting players are the design houses. Broadcom’s OpenAI deal and Marvell’s custom-chip and interconnect work with hyperscalers exemplify a broader trend of AI players turning to semi vendors to produce application-specific accelerators rather than buying off the shelf. And neoclouds like CoreWeave and Lambda build GPU-dense AI clouds in close partnership with Nvidia while also experimenting with AMD and custom chips. TensorWave’s all-AMD approach is a useful reminder that not every large-scale AI provider is bound to Nvidia — even if the biggest ones still are.

Not all hardware is the same

The whole landscape makes more sense once you separate workloads by type. Frontier training runs almost exclusively on general-purpose Nvidia hardware, and for good reasons. Developers need maximum flexibility, GPUs are more general than ASICs, and CUDA tooling cuts time-to-first-train dramatically. Performance per chip for these experimental runs is also still often best on Nvidia’s newest architectures. When you’re spending hundreds of millions on a training run, you don’t gamble on immature silicon.

Steady-state, high-volume inference is the opposite case. Once a model architecture stabilizes it becomes worthwhile to hard-optimize silicon around it. That’s where the 40–65% TCO advantage of custom ASICs matters most, because the savings on power and cooling compound across billions of daily queries. This is precisely why TPUs, MTIA and Trainium dominate internal hyperscaler workloads while Nvidia holds the external enterprise business. Internally, hyperscalers control the full stack and can mandate their own chips. Externally, customers demand compatibility with standard open-source frameworks that are optimized for Nvidia, and enterprises generally prefer an ecosystem their teams already know.

The real bottleneck, though, is software. Custom chips can match or beat Nvidia on raw performance and cost for targeted scenarios, but compilers, debuggers, profilers and high-level frameworks all have to be adapted to each architecture, and developer familiarity with CUDA means any non-Nvidia environment carries transition friction. Hyperscalers are closing the gap with in-house compiler teams and by wiring their chips into PyTorch, JAX and TensorFlow through cloud-specific backends. But it’s slow, and it’s why most commentators expect Nvidia’s ecosystem to stay the default for external workloads for several more years.

So how dependent are hyperscalers on Nvidia, really? Still very — for frontier training, for external cloud customers, and anywhere CUDA compatibility is non-negotiable. The most credible mid-term read is that they significantly reduce Nvidia dependence for inference, and modestly for training, by 2028–2030. Nvidia’s margins probably compress somewhat as alternatives mature. But its volume, its ecosystem, and its role at the experimental edge of the stack stay large. The hyperscalers have stopped being purely Nvidia’s customers and started becoming its architectural rivals.