Count them. In the last five weeks:
- Meta and Qualcomm announced Dragonfly, a data-center CPU targeted at Meta’s own inference workloads.
- OpenAI and Broadcom unveiled Jalapeño, an inference-tuned chip built specifically for OpenAI’s model shape.
- Google committed 3.5 gigawatts of Broadcom-fabbed next-generation TPUs to Anthropic, siting the vast majority in the US.
- TechCrunch reported this week that Anthropic is in early discussions with Samsung about a custom chip of its own.
Nobody has run a press event with a tagline like “we are divorcing Nvidia.” Nobody needs to. The pattern is doing the announcing.
Why now
The uncomfortable arithmetic is that Nvidia’s margins are the labs’ cost of goods sold. When Anthropic prices Claude Sonnet 5 at two dollars per million input tokens and ten dollars per million output through August, every one of those dollars is negotiating with a Nvidia SKU somewhere upstream. The gross margin on H200-generation silicon has been talked to death and does not need to be litigated again here. The relevant fact is that the frontier labs’ revenue is now big enough, and their unit economics tight enough, that squeezing that gross margin becomes a strategic priority instead of a spreadsheet exercise.
Anthropic’s own disclosure this week put run-rate revenue above $30 billion, up from $9 billion at the end of last year. Over a thousand business customers now spend seven figures a year with them. At that scale, inference cost per token is not a rounding-error line item. It is the difference between a service that funds compute buildout and one that funds Jensen Huang’s next keynote.
The other force is workload specificity. Nvidia’s H-series and next-gen Rubin parts are extraordinary general-purpose accelerators, and that generality is what made the industry possible. But once a lab has settled on a model architecture and knows exactly what its attention pattern, KV-cache shape, and quantization tolerance look like, a chip tuned to that specific shape can beat a general-purpose part on performance-per-watt by wide margins. The custom-silicon pitch has always been “you know what you’re running, we can build the exact thing.” What changed in 2026 is that the labs finally know what they are running for long enough to justify the tape-out.
What each deal is actually buying
Not all custom silicon plays the same role.
Meta and Qualcomm’s Dragonfly is a data-center CPU story. Meta is not walking away from GPUs. It is trying to own the general-purpose infrastructure layer around the GPU, which is where an enormous amount of inference latency and cost actually hides. Whoever owns the host CPU owns a lot of the workload-orchestration story that AWS and Azure would otherwise collect margin on.
OpenAI’s Jalapeño is inference-first. That is the correct read of a lab whose revenue is increasingly gated by how cheaply it can serve GPT-class calls to enterprise customers. Frontier training still runs on Nvidia or on Microsoft’s Nvidia. Serving is where the money bleeds out, and serving is where a custom part gets to shine.
Google’s TPU roadmap, extended through Broadcom to Anthropic, is the closest thing to a real full-stack alternative. TPUs have been running Google’s own workloads for years and have real production maturity that other custom parts do not. The Anthropic deal converts a decade of internal Google infrastructure work into an external product line.
Anthropic and Samsung, if it happens, would be the newest and least-defined of the four. Samsung foundry has real spare capacity that TSMC does not, and Samsung has been closing the process-node gap. A US frontier lab running an inference part on Samsung silicon would also be a subtle geopolitical hedge, given that TSMC represents the industry’s single point of failure at advanced nodes.
What it does not mean
Nvidia does not lose in 2026. Nvidia probably does not lose in 2027 either. Every deal here is either committed to future silicon that has not shipped, currently second-source rather than first-source, or explicitly framed as complementary to a diversified stack. Nobody who is a serious buyer of frontier compute is genuinely walking away from H200s or Rubin next year.
What is happening is subtler and more structural. The frontier labs are getting comfortable with the idea that they are, functionally, silicon companies. Not chip designers in the Nvidia sense, but customers with enough scale and workload specificity to insist on parts built to their spec. That is the same posture that turned Apple from a PC vendor into a silicon-vertically-integrated platform. The labs are ten years behind Apple on that curve. They are moving fast to close it.
The two-year read
If you want the least-controversial forecast for how this shakes out by 2028, it looks like this:
- Nvidia continues to own frontier training for the foreseeable future. The software moat is real. The interconnect story is real. Rubin and its successors will be extremely good.
- Serving splits. A meaningful fraction of high-volume inference migrates onto custom silicon at each of the top labs, because the labs cannot afford not to move it. What percentage is anyone’s guess. Fifty is aggressive. Twenty is conservative. Somewhere in between is the shape.
- Broadcom becomes the default design partner for labs that do not want to build their own chip team from zero. It already effectively is that partner for OpenAI, Google, and now the Google-Broadcom slice of Anthropic. It is not clear why Meta went to Qualcomm on Dragonfly instead of Broadcom, and that itself is worth watching.
- Samsung foundry re-emerges as a real second source at advanced nodes. If the Anthropic conversation matures, that story accelerates.
- TSMC continues to be the industry’s actual single point of failure, and the second Taiwan headline is going to move a lot of pricing.
The unlovely truth in all of this is that the frontier labs have spent three years telling everyone that intelligence is the product and infrastructure is a means to an end. The behavior over the last five weeks says something different. The infrastructure is not a means to an end. It is the balance sheet. And nobody is going to keep handing Nvidia a healthy fraction of their gross margin forever, no matter how good the software stack is.