On April 6, Broadcom filed an 8-K with the SEC confirming two agreements that will reshape how artificial intelligence is built and priced for the next decade. The first is a long-term partnership with Google to co-develop custom AI chips through 2031, covering the TPU v7 (codenamed Ironwood), v8, and v9 series. The second is an expanded collaboration with Anthropic that will give the Claude maker access to approximately 3.5 gigawatts of TPU-based compute capacity beginning in 2027.
Mizuho analysts estimate the Anthropic relationship alone could generate $21 billion in AI revenue for Broadcom in 2026, rising to $42 billion in 2027. Those numbers would make Anthropic one of the largest single customers in semiconductor history.
But the filing contained a sentence that enterprise AI buyers should read carefully: "The consumption of such expanded AI compute capacity by Anthropic is dependent on Anthropic's continued commercial success." The 3.5-gigawatt figure is conditional, not guaranteed. And the conditions under which it becomes real — or doesn't — will determine the cost, availability, and strategic independence of the AI capabilities your enterprise depends on.
This is not a chip deal. It is the opening move in a structural realignment of the AI infrastructure stack — one where the companies building the most capable AI models are also securing control of the silicon that runs them.
The Numbers That Changed Everything
To understand why this deal matters, start with Anthropic's trajectory.
Anthropic's annualized revenue crossed $30 billion in April 2026, surpassing OpenAI's $25 billion for the first time. Business customers spending over $1 million per year doubled from 500 to over 1,000 in under two months following the company's Series G raise. Anthropic CEO Dario Amodei has committed $50 billion to American computing infrastructure, with the majority of new compute sited in the United States.
Now look at the infrastructure required to sustain that growth. According to Broadcom CEO Hock Tan, Anthropic consumed roughly one gigawatt of compute in 2026. The new commitment triples that capacity before the year is out. Anthropic plans to operate as many as one million TPUs to run and expand its Claude model family, citing major cost-to-performance gains over general-purpose GPUs.
The financial math is straightforward: at a $30 billion run rate, Anthropic needs infrastructure that costs tens of billions per year. Building that infrastructure on NVIDIA GPUs at $60,000 to $70,000 per GB200 accelerator is one path. Building it on custom Broadcom-manufactured TPUs at an estimated 44 percent lower total cost of ownership is another.
Anthropic chose the second path. And that choice has consequences that cascade directly into every enterprise AI procurement decision in 2026 and beyond.
Inside the Ironwood Advantage
Google's seventh-generation TPU — Ironwood — is the chip at the center of this deal, and its specifications explain why Anthropic made this bet.
Each Ironwood chip delivers 4,614 FP8 teraflops of performance with 192 GB of HBM3E memory and 7.37 terabytes per second of memory bandwidth. In raw compute, that puts it fractionally ahead of NVIDIA's B200 at 4,500 teraflops and just below the GB200 and GB300 at 5,000 teraflops.
But the comparison at the chip level misses the point. The real advantage is at scale.
Ironwood is designed for massive deployment. Google offers the chips in pods ranging from 256 to 9,216 chips. At the high end, an Ironwood pod delivers 42.5 FP8 exaflops of compute for both training and inference — a figure that dwarfs NVIDIA's GB300 NVL72 system at 0.36 exaflops. The difference is not incremental. It is two orders of magnitude.
For the technical audience, the interconnect architecture is what makes this possible. Broadcom provides the 9.6 Tbps SerDes interconnects that stitch Ironwood chips together across Google's custom 3D Torus network fabric. This is not off-the-shelf networking. It is a vertically integrated compute stack where the chip, the interconnect, and the network topology are co-designed — and it is manufactured on 2nm and 3nm process technology.
For the business audience, the translation is simple: Google and Broadcom have built a compute architecture that is purpose-built for AI workloads, runs at dramatically lower cost per computation than general-purpose GPU systems, and scales to sizes that NVIDIA's current product lineup cannot match in a single coherent cluster.
SemiAnalysis estimates that Google cuts its cost per computation by 62 percent relative to NVIDIA with its internal TPU workloads. Midjourney reported that migrating from NVIDIA GPUs to Google TPUs reduced monthly compute costs from $2.1 million to $700,000 — a 65 percent reduction. Google's own benchmarks claim TPU v6e offers up to four times better price-performance for inference workloads than NVIDIA H100, with 67 percent less power consumption.
These are the economics driving Anthropic's decision. When you are spending $30 billion a year and growing, a 44 to 62 percent cost reduction on your largest expense line is not a procurement optimization. It is a survival strategy.
The Great Silicon Pivot
Anthropic is not alone. Every major AI lab and hyperscaler is now building or commissioning custom silicon.
Amazon's Trainium 3 targets 30 to 40 percent better price-performance than comparable GPUs, with the company projecting "tens of billions" in capital expenditure savings and several hundred basis points of operating margin advantage. Microsoft's Maia 200 is designed to reduce Azure's dependence on NVIDIA for internal AI workloads. Meta's MTIA chip is optimized for recommendation and ranking models at scale.
The aggregate trajectory is unambiguous. Custom AI accelerator shipments are projected to triple by 2027 compared to 2024. The custom ASIC market is growing at 44.6 percent annually, while GPU-based solutions grow at 16.1 percent. By 2028, ASIC shipments are forecast to surpass GPU shipments for the first time in semiconductor history.
NVIDIA sees this clearly. Its revenue share peaked near 87 percent in 2024 and is projected to decline to 75 percent by 2026. Analysts project its inference market share could fall from over 90 percent to 20 to 30 percent by 2028. NVIDIA's response includes a $2 billion investment in Marvell, Broadcom's primary competitor in the custom ASIC space, and the NVLink Fusion initiative — an attempt to create an ecosystem lock-in that makes custom chips interoperable with NVIDIA's networking stack rather than replacing it entirely.
But the physics of the market favor specialization. Custom ASICs designed for specific workloads — inference, training, embedding generation — can optimize transistor allocation, memory hierarchy, and power consumption in ways that general-purpose GPUs cannot. When inference now represents two-thirds of all AI compute, the incentive to build chips that are purpose-optimized for inference rather than general-purpose training is enormous.
Broadcom sits at the center of this transition. It commands approximately 72 percent of the data center XPU market by revenue, generating roughly $5.1 billion per quarter in custom accelerator revenue. Its AI semiconductor revenue reached $8.4 billion in fiscal Q1 2026, up 106 percent year-over-year. Management has guided toward exceeding $100 billion in cumulative AI semiconductor revenue by fiscal year 2027.
What This Means for Enterprise AI Buyers
If your organization consumes AI through cloud APIs — and in 2026, most do — the custom silicon pivot affects your costs, your vendor strategy, and your negotiating position in three specific ways.
First, inference costs are heading down — but unevenly. The 44 to 65 percent cost advantage of custom silicon over NVIDIA GPUs flows first to the AI labs and hyperscalers that own the chips. Whether and when those savings reach enterprise API consumers depends entirely on competitive dynamics.
Anthropic running Claude on purpose-built TPUs at 62 percent lower cost per computation than NVIDIA gives Anthropic enormous pricing flexibility. It can use that margin to undercut OpenAI on per-token pricing, invest in capability development, or simply take the profit. The enterprise buyer's leverage depends on having credible alternatives. If Claude is running on Google infrastructure and GPT is running on Microsoft infrastructure, the ability to arbitrage between them becomes the enterprise's primary pricing lever.
The practical implication: enterprises that lock into single-vendor AI contracts in 2026 will have less pricing leverage as custom silicon widens the cost gap between providers. Multi-model, multi-vendor strategies become more valuable precisely because the underlying cost structures are diverging.
Second, Anthropic's Google Cloud dependency creates a structural vendor tension. As Futurum Group's analysis noted: "At what point does Anthropic's Google Cloud reliance become a liability in enterprise deals where Microsoft Azure is the incumbent?"
This is not theoretical. If your enterprise is standardized on Azure and you want to deploy Claude at scale, the compute is running on Google-manufactured TPUs in Google-operated data centers. Your data governance, your compliance framework, and your procurement relationships are routed through a supply chain that your primary cloud vendor does not control.
For CIOs managing multi-cloud strategies, this creates a new variable in vendor evaluation. The model you want may not run on the cloud you already pay for — at least not at the same cost or performance level. Anthropic offers Claude through Amazon Bedrock and its own API, but the TPU-optimized deployment runs on Google Cloud. The performance and cost parity across deployment targets is not guaranteed.
Third, the timeline matters. The Broadcom deal delivers 3.5 gigawatts of new TPU capacity starting in 2027, with the TPU v8 and v9 roadmap extending to 2031. Custom silicon advantages take 12 to 24 months to mature at production scale.
For enterprises making AI infrastructure decisions today, this means the cost landscape of 2027 will look materially different from 2026. Signing long-term committed-use contracts at today's pricing may lock in costs that look expensive within 18 months. Conversely, waiting for custom silicon savings to flow through to enterprise pricing means deferring AI deployment during a period when competitors may be building advantages.
The CFO's calculation is a classic timing problem: commit now at higher costs to capture first-mover advantage, or wait for the custom silicon wave to reduce pricing at the risk of falling behind.
The Control Plane Shifts Upstream
The deeper strategic implication of the Broadcom deal is that frontier AI companies are moving upstream into infrastructure.
When Anthropic secures 3.5 gigawatts of dedicated compute capacity — enough to power a mid-sized city — it is not simply buying chips. It is building a vertically integrated operation where the model, the training infrastructure, the inference stack, and the underlying silicon are all controlled or co-designed by the same entity.
This is the pattern that Apple pioneered with its M-series chips: vertical integration from silicon to software, with the explicit goal of controlling the full stack to optimize performance and cost in ways that horizontal competitors cannot match.
For enterprise buyers, vertical integration by your AI supplier means two things. On the positive side, it means lower costs and better performance as the stack is optimized end-to-end. On the negative side, it means less leverage. When Anthropic controls the silicon, the training infrastructure, and the API, the traditional enterprise procurement tactic of playing vendors against each other at the hardware layer no longer applies.
The countervailing force is competition. Anthropic on Google TPUs competes with OpenAI on custom Broadcom ASICs (Citi estimates that deal at $100 to $200 billion over multiple years) and with Amazon's Trainium-powered internal models. The enterprise buyer's leverage comes from model-layer competition, not infrastructure-layer competition.
This is a fundamental shift. For the past two decades, enterprise IT procurement has relied on commodity hardware creating a competitive market at the infrastructure layer. In the AI era, the infrastructure layer is being absorbed by the application providers. The competition moves up to the model and capability layer.
Enterprise procurement teams need to update their frameworks accordingly. The question is no longer "which cloud has the best GPU pricing." It is "which AI provider delivers the best capability per dollar, and how do I maintain switching optionality as their infrastructure becomes increasingly proprietary."
What To Do Now
For enterprise AI leaders navigating the custom silicon transition:
Audit your AI cost structure. Understand what percentage of your AI spend goes to compute versus model access versus engineering. As custom silicon drives down the compute component, the relative weight of other cost drivers — data preparation, integration, governance — increases. Budget allocations should reflect this shift.
Maintain multi-vendor optionality. The diverging infrastructure stacks of Anthropic (Google TPUs), OpenAI (Broadcom ASICs), and Amazon (Trainium) make single-vendor lock-in increasingly risky. Architect your AI applications with abstraction layers that allow model and provider switching.
Negotiate shorter contract terms. With custom silicon delivering 44 to 62 percent cost reductions that will take 12 to 24 months to flow through to enterprise pricing, long-term committed-use agreements signed in 2026 may not reflect the cost reality of 2027. Push for shorter commitments or pricing renegotiation clauses tied to infrastructure cost indices.
Watch the inference pricing war. Custom silicon's biggest impact will be on inference costs, which now represent two-thirds of all AI compute. When Anthropic can run inference at 62 percent lower cost and Amazon offers Trainium at 30 to 40 percent savings, the providers with the lowest inference costs will win enterprise workloads that run continuously. Price your AI workloads accordingly.
The custom silicon pivot is the most consequential infrastructure shift in enterprise AI since the move to cloud computing. The companies that built the most capable AI models are now building the chips that run them. The cost of intelligence is about to drop — and the enterprise leaders who position for that drop will capture the value.
Want to calculate your own AI ROI? Try our AI ROI Calculator — takes 60 seconds and shows projected savings, payback period, and 3-year ROI.
