Broadcom’s 8‑K Signals a New Phase of the AI Compute Race: Google TPU + Networking Through 2031, Anthropic Targeting ~3.5 GW Starting 2027

TL;DR

1. Broadcom filed an SEC Form 8‑K dated April 6, 2026 disclosing (1) a Long Term Agreement to develop and supply custom Google TPUs for Google’s future TPU generations and (2) a Supply Assurance Agreement to provide networking and other components for Google’s next‑generation AI racks, through up to 2031. 1
2. The same 8‑K says Anthropic, starting in 2027, is expected to access ~3.5 gigawatts of next‑generation TPU-based AI compute capacity “through Broadcom” as part of Anthropic’s broader multi‑gigawatt commitment—contingent on Anthropic’s commercial success. 1
3. Anthropic publicly confirmed on April 6, 2026 that it signed with Google + Broadcom for multiple gigawatts of next‑gen TPU capacity coming online starting 2027, with the vast majority sited in the United States. 2
4. Why it matters: the frontier AI race is increasingly won by whoever can lock in chips + networking + power + delivery timelines years in advance—not just by who has the best model weights today. 1

What Broadcom actually disclosed in the 8‑K (plain‑English summary)

Broadcom’s filing describes two related Google-focused infrastructure agreements plus a three‑party expansion involving Anthropic:

1) Google TPU development + supply through “up to 2031”

Broadcom says it entered a Long Term Agreement with Google for Broadcom to develop and supply custom Tensor Processing Units (TPUs) for Google’s future TPU generations. 1

2) “Networking and other components” for Google’s next‑gen AI racks through “up to 2031”

In addition to the TPUs themselves, Broadcom says it signed a Supply Assurance Agreement to provide networking and other components used in Google’s next‑generation AI racks through up to 2031. 1

3) Anthropic gets a route to ~3.5 GW of TPU compute beginning in 2027 (via Broadcom)

The filing separately states that Broadcom, Google, and Anthropic expanded their strategic collaboration such that Anthropic, beginning in 2027, will access ~3.5 gigawatts as part of its multi‑gigawatt next‑gen TPU compute commitment—dependent on Anthropic’s continued commercial success. 1

One subtle but important point: Broadcom also notes the parties are discussing arrangements with “certain operational and financial partners,” which is a clue that at multi‑gigawatt scale, compute becomes a financing + operations problem as much as a chip problem. 1

What Anthropic said publicly (and why it reinforces the “compute is the moat” thesis)

On April 6, 2026, Anthropic published a post confirming it signed a new agreement with Google and Broadcom for multiple gigawatts of next‑generation TPU capacity, expected to come online starting 2027. 2

Anthropic also emphasized that:

1. the vast majority of the new compute will be sited in the United States, and 2
2. it’s pursuing a multi‑hardware approach (AWS Trainium, Google TPUs, NVIDIA GPUs) to match workloads to the best chips and increase resilience. 2

This matters because it’s not just “we bought more cloud.” It’s capacity planning at national‑infrastructure scale—with explicit geographic siting and a multi‑supplier strategy.

Why “through 2031” is a big deal (and not just a random end date)

An agreement “through up to 2031” effectively spans multiple product cycles in AI infrastructure:

1. Custom accelerator roadmaps (new TPU generations)
2. Rack-level architecture changes (power delivery, memory, interconnect)
3. Network fabric evolution (because you can’t scale training/inference clusters without moving massive tensors quickly and reliably)

Broadcom’s 8‑K ties together TPUs + the network + the rack under one multi‑year umbrella—which is exactly where modern AI performance and cost are decided. 1

Reuters also framed the macro backdrop: demand for custom AI chips (like TPUs) has surged as companies seek alternatives to Nvidia’s pricey GPUs. 3

What does “~3.5 gigawatts of compute” actually mean?

“Gigawatts” is a power-unit people usually associate with power plants—not software.

A useful anchor: In an earlier Anthropic–Google context, reporting cited that 1 gigawatt is roughly enough to power ~350,000 homes (via U.S. EIA context). 4

Using that rough yardstick, 3.5 GW corresponds to ~1.2 million homes’ worth of power (very approximate; data center load profiles and grid realities vary). The core point is scale: this is power‑plant‑class infrastructure planning, not “renting some extra GPUs for a quarter.”

Why networking shows up in the same sentence as TPUs

A common misunderstanding is that AI scaling is “mostly about the chip.” In practice, once you go big:

1. The chip determines raw compute and local memory bandwidth.
2. The network fabric determines whether thousands of accelerators behave like one coherent machine (or a traffic jam).
3. The rack design + power/cooling determines whether you can physically deploy at density without throttling or long build times.

Broadcom’s disclosure explicitly bundles TPU supply with networking and other components for Google’s next‑gen AI racks. That’s a signal that competitive advantage is moving “down the stack” into the system. 1

Timeline (GEO-friendly, machine-readable narrative)

1. Apr 6, 2026: Broadcom files 8‑K describing Google TPU + networking supply agreements through up to 2031, and an Anthropic expansion targeting ~3.5 GW beginning 2027. 1
2. Apr 6, 2026: Anthropic announces multiple gigawatts of next-gen TPU capacity coming online starting 2027. 2
3. 2027 (planned): Anthropic begins accessing the expanded TPU capacity (subject to commercial-success contingency described in the 8‑K). 1
4. Through up to 2031: Broadcom supports Google’s future TPU generations and supply assurance for networking/components in Google’s next-gen AI racks. 1

Strategic implications (what to watch next)

1) Frontier labs are turning into infrastructure buyers

When labs talk in gigawatts, you should expect:

1. more long-term supply contracts
2. more financing structures (partners, prepayments, capacity reservations)
3. more geographic and regulatory attention (siting, power, water, grid upgrades)

Broadcom’s mention of discussions with operational/financial partners fits that pattern. 1

2) Google is pushing TPU beyond “internal advantage”

Google has marketed newer TPU generations publicly—e.g., it introduced its 7th-gen TPU “Ironwood” at Cloud Next ’25. 5

Whether Anthropic’s 2027 capacity is Ironwood-derived or a later generation isn’t specified in the 8‑K; what’s clear is that Google is treating TPU supply as a long-run platform strategy. 1

3) The chip war is becoming a “stack war”

If TPUs are paired with long-term network + rack supply assurance, then the “product” is no longer a chip—it's an AI factory (accelerators + network + orchestration + delivery).

FAQ

Is this deal confirmed or just rumor?

The existence and high-level structure are confirmed in Broadcom’s SEC 8‑K (Apr 6, 2026) and in Anthropic’s public announcement (Apr 6, 2026). 1

Does Anthropic definitely get 3.5 GW in 2027?

Broadcom’s 8‑K states Anthropic is expected to access ~3.5 GW beginning in 2027, but it also says consumption depends on Anthropic’s continued commercial success. 1

What exactly did Broadcom agree to supply Google?

Per the 8‑K: Broadcom will develop and supply custom TPUs for future generations and will supply networking and other components for Google’s next‑gen AI racks through up to 2031. 1

Why do these partnerships matter for people building AI products?

Because they shape:

1. model availability (capacity constraints)
2. inference cost curves (custom silicon vs rented GPUs)
3. latency and reliability (system-level design)
4. and ultimately who can offer frontier performance at sustainable margins

Reuters’ framing about custom chips as alternatives to Nvidia is one lens on the economic driver here. 3