Broadcom Reports AI Revenue Surge to $10.8 Billion in the Quarter and Aims for $100 Billion by 2027 with Custom Chip

Broadcom reported $22.2 billion in revenue for the second fiscal quarter of 2026, an increase of 48% year-on-year, with AI semiconductor revenue reaching $10.8 billion, a leap of 143% year-on-year. The results, disclosed on June 3 after market close, exceed the company's own guidance upper limit and expand the annual AI revenue forecast to $56 billion for fiscal year 2026, a rise of 180% compared to 2025. The target for fiscal year 2027 has now been set at over $100 billion.

Adjusted EBITDA hit a record of $15.2 billion, equivalent to 69% of quarterly revenue. Networking accounted for nearly 40% of AI revenue, with the company citing demand for custom accelerators and Tomahawk Ultra and Jericho4 switches in hyperscaler deployments. The guidance for the third quarter projects semiconductor AI revenue of $16 billion, an increase of over 200% year-on-year.

Who is Buying and How Much

In the earnings call, Hock Tan detailed the six main customers for custom chips: Google, Meta, OpenAI, Anthropic, and two still undisclosed publicly. Analysts from Bernstein and Morgan Stanley suggest Apple and ByteDance as likely candidates, but Broadcom has not confirmed this. Anthropic was the customer with clarified scale mentioned in the call: a contract for 1 gigawatt in TPUs for 2026, scaling up to 3 gigawatts in 2027. OpenAI is expected to receive its first custom chip in 2027, according to Tan, coinciding closely with the launch window of the Stargate Project with Oracle.

The announcement of Anthropic's figures provides insight into the IPO that the Claude manufacturer is preparing for October. The implied capital expenditure for 3 gigawatts of TPUs in a single year amounts to tens of billions of dollars, explaining why the company is pricing its IPO at an approximate valuation of $1 trillion and aiming for up to $60 billion in fundraising. Goldman Sachs and Morgan Stanley are leading this syndicate and are also the equity research firms that have aggressively covered Broadcom with a buy rating since 2024.

Revenue Concentration, and What Nvidia's Competitor Sees

Broadcom's custom ASIC represents a structural alternative to Nvidia's GPU in hyperscaler deployments. Nvidia still dominates the AI accelerator market share due to software margins and CUDA lock-in, but the total cost per watt in next-generation data centres has led Google, Meta, and Anthropic to allocate inference workloads to custom chips to free up Hopper and Blackwell GPUs for training. This relationship reaches an absolute figure in guidance: Broadcom's $56 billion AI revenue in 2026 is larger than the entire data centre revenue reported by AMD in 2025 ($12.6 billion) and equates to approximately 25% of Nvidia's projected data centre segment for the same fiscal year.

Geographic insights matter for those operating fabs. Broadcom does not manufacture chips; it relies on TSMC in Taiwan for N3 and N2 nodes and on Samsung Foundry in South Korea for part of the Tomahawk production. The $100 billion guidance for 2027 presumes preferential wafer allocation from TSMC, placing the Hsinchu manufacturer in a pivotal position among Nvidia, Apple, and Broadcom amidst competition. The new capacity projected for Phoenix, Arizona by TSMC is set to enter production in 2026, but with volumes that do not cover the total demand from the three competing entities.

For the CIO in London, Tokyo, and São Paulo

For the CIO negotiating Google TPU consumption, Meta MTIA via API, Claude via Bedrock, or directly with Anthropic in any of the three markets, Broadcom's figure is the clearest sign to date of how much of the hyperscaler's margin relies on custom chips rather than off-the-shelf GPUs. The practical consequence is the negotiation window: multi-year contracts signed in the next six months will price in the expectation of unit cost reductions that will only materialise in 2027, after Anthropic's 3 GW of TPUs are in production and OpenAI receives its first custom chip. Those signing before this window bear the legacy of GPU-based pricing.

The secondary effect appears in the capex line of hyperscalers. Google, Meta, and Microsoft exited the first quarter with a combined guidance nearing $300 billion in capex for 2026; Broadcom's figure evidences that most of this capex will go towards custom silicon rather than Nvidia licenses. Regional cloud operators in Europe, Japan, and Brazil, lacking access to proprietary custom chips, become increasingly dependent on the economics of off-the-shelf GPUs, losing marginal ground in inference economics.