For the past two years, the AI compute race has had essentially one "arms dealer" — NVIDIA. But over the past year, the winds have quietly shifted: Google is pushing its own TPUs from internal use to external supply, Anthropic signed a compute deal for up to one million TPUs, OpenAI and Broadcom struck a custom accelerator partnership of up to 10 gigawatts, and Amazon built a dedicated training cluster of hundreds of thousands of cards for Anthropic using its in-house Trainium. Cloud giants are moving from "buying chips" to "making — and even selling — chips," and a wave of custom ASICs is rewriting the competitive map of US-listed AI chips. This is more than a battle over technical routes: it will decide where hundreds of billions of dollars in capital spending flow in the years ahead, and it deserves a close read from every investor tracking the US AI compute trade.
Cloud Giants Go All-In on Making Chips: The Critical Leap from In-House Use to External Supply
Custom chips are nothing new. Google's TPU has iterated to its seventh generation, Ironwood, optimized for inference; Amazon's Trainium family and Microsoft's Maia have run on their respective clouds for years. The real change is the "external supply" step: Google Cloud announced it will open up to one million TPUs of compute capacity to Anthropic, with deliveries exceeding 1 gigawatt starting in 2026; Meta has reportedly been in talks to rent TPUs via Google Cloud to train its own models. Meanwhile, after landing gigawatt-scale deals with NVIDIA and then AMD, OpenAI signed a multi-year joint development program with Broadcom for custom accelerators, with the first systems expected to deploy from 2026. Amazon's Project Rainier has assembled nearly 500,000 Trainium2 chips dedicated to training Anthropic's next-generation models.
Taken together, these moves send one clear signal: leading model developers' hunger for compute has grown too big to rely entirely on any single chip supplier, and "multi-supplier plus in-house" has become the default strategy. The power to choose chips is partly shifting back from NVIDIA to the cloud giants themselves.
Why Cloud Giants Insist on Building Their Own Chips: Three Ledgers That Must Add Up
First: The Cost Ledger
Once AI entered its inference-dominated phase, procurement logic shifted from "buying the strongest cards" to "counting how many tokens each dollar and each watt delivers." The underutilization of general-purpose GPUs under fixed workloads gets amplified, while ASICs deeply optimized for one's own model architecture often open a clear gap in unit cost and energy efficiency. For giants with annual capex in the tens of billions of dollars, even a 20-30% saving is a huge number.
Second: The Bargaining-Power Ledger
Compute has become a strategic resource, and staking one's lifeline on a single supplier is too risky. In-house chips are both a technical fallback and leverage at the NVIDIA negotiating table — only with a second option is there room to push on pricing and supply priority.
Third: The Differentiation Ledger
As cloud offerings grow increasingly homogeneous, "integrated software and hardware" has become the new selling point: Google bundles TPUs with its Gemini and Vertex platforms, and Amazon embeds Trainium into Bedrock. In-house chips have become a moat that sets cloud providers apart from peers, while external supply turns that massive R&D outlay into a revenue stream as well.
A New Roster of "Shovel Sellers": The Custom ASIC Value Chain at a Glance
For US equity investors, the biggest significance of this wave is that AI chip money no longer flows to NVIDIA alone — a new value chain is surfacing.
- Design services and IP licensing: Broadcom is among the biggest winners; its custom XPU business already serves multiple hyperscale customers, and with the OpenAI deal layered on, custom chips and networking chips now drive AI revenue on twin engines. Marvell is deeply tied to custom projects such as Amazon's Trainium, with related orders growing fast.
- Manufacturing and packaging: Whoever designs the chips, advanced process nodes and advanced packaging almost inevitably run through TSMC; 3nm capacity and CoWoS packaging remain persistently oversubscribed, giving the foundry segment highly certain momentum.
- Interconnects and supporting parts: The more custom chips, the more complex chip-to-chip and cluster-level interconnects become; interconnect vendors such as Credo and Astera Labs, along with optical modules and switch chips, benefit in step.
- Memory: HBM is standard equipment on every AI chip, and HBM orders at SK hynix, Samsung, and Micron climb as custom chip volumes ramp.
In one sentence: the rise of custom ASICs is pushing the allocation of AI compute value from "one dominant player" to "gains across the board," letting design, manufacturing, packaging, interconnect, and memory each share the incremental pie.
Can NVIDIA's Moat Really Be Shaken?
As custom chips heat up, the question markets care about most is whether NVIDIA's position is wobbling. For now, the answer is closer to "eroded, but hard to dislodge." NVIDIA's advantages — the CUDA ecosystem, a once-a-year product cadence, and full-rack delivery powered by NVLink rack-scale systems — remain hard to replace for most customers. What custom ASICs capture is mainly incremental workloads with highly predictable patterns among hyperscale customers, concentrated on the inference side; the long tail of fast-iterating models and diverse demands will stay in the general-purpose GPU camp. The more realistic picture: the future AI chip market will run on dual tracks of "general-purpose GPUs plus custom ASICs." NVIDIA's growth rate may be diluted, but its absolute scale keeps expanding, and its valuation narrative shifts from "the only choice" to "one of the best choices" — precisely the core variable the market now has to reprice.
Investment View: Three Themes and Three Risks
At the investment level, three themes stand out relatively clearly in the custom ASIC wave: first, chip design services leaders holding major customer orders; second, manufacturing momentum born of scarce advanced-node and advanced-packaging capacity; third, "must-have supporting" segments such as interconnects, optical modules, and HBM.
The risks are just as real: first, custom chips carry long tape-out cycles and steep failure costs — once a client project slips, earnings upside can turn into earnings pressure overnight; second, customer concentration is high, making bargaining power and order volatility risks prominent; third, sector valuations already price in elevated expectations, and if AI capex growth slows, the pullback could be just as amplified.
For everyday investors, rather than betting on a single name, it makes more sense to put this value chain on a watchlist and track three kinds of signals: shifts in cloud providers' capex guidance, the disclosure pace of custom chip projects, and utilization rates at foundries such as TSMC. In the second half of the AI compute race, the contest is no longer just whose chips are fastest, but who can drive the total cost per unit of compute lowest — and this efficiency battle has only just begun. This article is industry analysis only and does not constitute any investment advice.
