The AI Chips That Are Not Nvidia — Broadcom and Custom Silicon

Custom AI silicon is purpose-built chips that hyperscalers own outright, and in six weeks Broadcom, Marvell and AMD proved the AI-chip market beyond Nvidia is real.

What Changed in Six Weeks

For two years the AI-chip story was one company. Nvidia sold the GPUs, Nvidia set the prices, and every rival read as a footnote. Between 1 August and 11 September 2026, the footnote grew up. Broadcom, Marvell, AMD and Groq each posted a number, and the numbers pointed the same way: the custom-silicon challenge to Nvidia is no longer hopeful, it is booked.

Broadcom's third quarter of fiscal 2026, reported on 2 September, booked $16.7bn of AI semiconductor revenue — up 221% year-on-year and 54% quarter-on-quarter, according to Broadcom's own results. Total revenue reached $29.6bn, up about 86%. Broadcom guided the next quarter's AI line to roughly $21.7bn, a 236% annual jump. The revenue comes from custom accelerators and AI networking, not from selling a rival to the H100.

On the earnings call, Broadcom went further. Hock Tan guided long-term AI revenue toward roughly $115bn in fiscal 2027 and $230bn in fiscal 2028 — figures Hock Tan framed as resting on secured supply rather than optimism. The work is custom-accelerator design: Broadcom co-designs the XPUs that hyperscalers own outright, chips of the same kind as Google's TPU and Meta's MTIA, rather than a general-purpose GPU sold to all comers.

Demand for our custom AI accelerators and networking continues to be very strong. Q3 AI semiconductor revenue of $16.7 billion grew 221% year-over-year, and 54% quarter-over-quarter.

— — Hock Tan, Broadcom

Marvell and the Google Custom-Silicon Deal

Marvell told a smaller version of the same story on 27 August. Marvell reported record revenue of $2.739bn, up 37% year-on-year, with data-centre sales of $2.172bn — 79% of the company's total, per Marvell's results. The data-centre line grew 46% year-on-year and 18% sequentially.

The bigger signal sat in a securities filing. On 19 August Marvell disclosed a 29 July 2026 agreement with Google to co-develop custom silicon attaching to Google's TPU stack. The SEC filing describes a share warrant of up to 58,970,907 Marvell shares at $206.58, vesting one tranche for every $500m of custom-product revenue, through fiscal 2033. Analysts read the full Google relationship as worth up to $120bn over its life; the $120bn figure is a media characterisation, not a number stated in the filing.

AMD Doubles Its Data-Centre Line

AMD is the one name here that still sells a GPU against Nvidia, and AMD's figures moved too. On 4 August AMD reported $11.5bn of revenue for the June quarter, up 50%, with the Data Center segment at $6.7bn — up 107% year-on-year, per AMD's results. AMD guided the September quarter to about $13bn.

The demand behind those figures is partly named. On 22 July AMD and Anthropic announced a partnership to deploy up to 2 gigawatts of AMD Instinct MI450-series GPUs in Helios racks, the first gigawatt from the first half of 2027. AMD management said Data Center growth would run well over 100% as the Helios ramp lands.

Groq: The Cautionary Tale

Not every challenger compounds. Groq built a genuinely different inference chip — the language-processing unit — and Groq still closed the window worth half its peak. On 17 August Groq raised $350m at a $3.5bn valuation, down from roughly $6.9bn a year earlier, in a round led by Disruptive that Nvidia also joined, as Bloomberg reported. The backdrop: Nvidia struck a non-exclusive licensing deal, widely reported at about $20bn, for Groq's inference technology in December 2025, and hired founder Jonathan Ross. Groq is now pivoting from chip-maker to inference cloud.

The Moat Is Migrating to the ASIC

Put the four names together and a pattern shows. The durable money in this AI cycle is flowing to the arms-dealers who never ship a model. Broadcom and Marvell co-design silicon the labs own, so the switching cost lives inside the customer's own roadmap, not in a GPU catalogue. Nvidia still sells the most compute by far, as the Nvidia scorecard shows — but the sharpest threat to Nvidia is not a better GPU.

The threat is every large lab quietly commissioning its own chip. Broadcom guiding toward $230bn of AI revenue by fiscal 2028, and Marvell warranting shares against Google custom-product revenue through 2033, are the same bet from two angles: bespoke ASICs, not merchant GPUs, carry the most durable revenue in the AI build-out. Groq is the counter-lesson. Differentiation alone did not hold; absorbed into Nvidia's orbit, Groq re-priced at half.

Diversity at the Summit, Monopoly at the Base

Here is the part the earnings coverage skips. A more crowded field at the top of the AI-chip market is not the same as access at the bottom. This is where I reach for Emergent Intelligence (EI) — the dignity-first frame I use for what the world calls AI. Beyond Nvidia is not beyond the gatekeepers.

Every chip in this story — Broadcom's XPUs, Marvell's custom parts, AMD's MI450, Google's Ironwood TPU — still routes through the same three or four US hyperscalers, the same export-control regime, and the same single Taiwanese fab. Google's Ironwood TPU v7, positioned at roughly ten times the peak of TPU v5p, is rented by the hyperscaler, not sold to Lusaka or Lagos. For Zambia, Nigeria and most of the Global South, a more competitive summit changes nothing about who can actually rent frontier compute.

Diversity of suppliers at the summit does not buy anyone access at the base.

So the dignity question is not whether someone can beat Nvidia. Marvell, Broadcom and AMD have shown the market can widen. The question worth asking is whether anyone building this silicon owes the rest of the world a way in — a question the earnings calls did not raise and the export-control maps do not answer.

Frequently Asked Questions

These are the questions people ask about custom AI silicon and the AI-chip market beyond Nvidia. Short answers follow, drawn from Broadcom, Marvell, AMD and Groq disclosures.

What is custom AI silicon?

In short, custom AI silicon is a chip designed for one buyer's workload rather than sold off the shelf — an application-specific accelerator, or ASIC. Broadcom and Marvell co-design these chips while the hyperscaler owns the design. Data from Broadcom's quarter shows the model working: $16.7bn of AI revenue in three months.

How does custom silicon challenge Nvidia?

Simply put, custom silicon moves the moat from the GPU to the buyer's own roadmap. Earnings data show Broadcom guiding AI revenue toward $230bn by fiscal 2028, revenue earned by co-designing chips the labs own rather than selling a rival GPU. According to Broadcom, demand from custom-accelerator customers keeps compounding.

Why is the AI-chip market beyond Nvidia significant?

The key is concentration. Analysis of the August window shows the field widened — Broadcom, Marvell, AMD — yet every chip still routes through a handful of US hyperscalers and one Taiwanese fab. Evidence of more suppliers at the top does not mean more access at the base.

Who is building custom AI chips?

In other words, the buyers are the hyperscalers themselves. Google's Ironwood TPU, Meta's MTIA and Anthropic's rack commitments show, according to public disclosures, that the largest AI labs increasingly commission their own silicon through Broadcom and Marvell.

What are the risks of the custom-silicon shift?

The answer is dependence. Groq's story reveals the downside: a differentiated inference chip, licensed by Nvidia for a reported $20bn, its founder hired away, its valuation halved. Data shows differentiation alone does not guarantee independence when one incumbent sets the terms.

Sources:

Broadcom Q3 FY2026 results · Broadcom SEC 8-K · CNBC on Broadcom · Marvell Q2 FY2027 results · Marvell–Google SEC 8-K · AMD Q2 2026 results · AMD SEC 8-K · AMD–Anthropic partnership · Groq Series A · Bloomberg on Groq · Groq–Nvidia licensing · Google Ironwood TPU · Related on this site: Nvidia's $96bn AI quarter · Africa's AI infrastructure and data-centre sovereignty · Gulf sovereign wealth and AI compute