Can Broadcom save the semiconductor industry?

👆Click the blue text to follow us

Broadcom

Broadcom missed expectations last quarter and indirectly kicked off a de-leveraging wave of epic proportions in the semiconductor industry. After three months of sluggishness, can it revive morale and lead a sector rebound this time?

The financial report itself was lackluster: revenue and profit were in line with expectations, AI semiconductors beat expectations but were still not in the optimistic range, and the market responded by selling off immediately. However, during the earnings call, CEO Hock Tan revealed some explosive information, which is much more worth noting than the numbers themselves. So, let's skip the boring earnings and get right to the call highlights.

Regarding future AI revenue growth, management expects AI semiconductor revenues to reach $58 billion in fiscal year 2026, slightly above expectations, nearly double to $115 billion in 2027, which is below expectations, and double again to $230 billion in 2028, greatly exceeding expectations and even 5% higher than the most optimistic forecasts. In other words, the management, modeling after Jensen Huang, provided very clear long-term guidance, which to a large extent soothed market concerns.

As for profit margins, management stated that as the proportion of custom accelerators in revenue increases, and as more memory is incorporated into chips, the overall gross margin will be affected by changes in product mix. The Q4 gross margin is expected to be 73%, below market expectations. However, since revenue growth will outpace expenses, operating margins can remain stable.

Jason believes that giving long-term revenue guidance is likely an imitation of Nvidia’s approach—using a bright, clear number to reassure the market. But because they gave no long-term guidance for gross margins, perhaps the 73% in the next quarter may not be the bottom for margins.

On key customer collaborations, management said they are working closely with Google, and through long-term agreements have settled on ongoing multi-generation TPU and AI networking product partnerships, expecting to deliver tens of billions of dollars in TPUs to Google annually.

Regarding collaboration with Anthropic, management said Anthropic is expected to add 5 GW of TPU computing power in 2027 and a further 10 GW in 2028. The management emphasized that these 10 GW are exclusively for Anthropic and do not include Google’s demand. Even if Google’s business continues to grow, Anthropic is still expected to become Broadcom’s largest custom client by 2027.

Jason believes this basically confirms two things. First, the proportion of Google's in-house contribution is declining. Second, Anthropic may build its own data center as early as next year.

Next, regarding cooperation with OpenAI, management said a hard-to-pronounce J-something accelerator is progressing according to a 2027 plan for 1.3 GW. By 2028, it is expected to exceed 5 GW, becoming Broadcom’s second largest custom client. Both parties are also developing second and third generation products, and the partnership will span multiple chip generations.

As for Meta, management said that mass production and delivery will begin from next quarter. Over about five quarters until the end of 2027, three generations of products are planned for delivery to Meta. By 2028, related deployments are expected to reach 3 GW.

The total demand from these clients over the next two years is about 30 GW. However, Broadcom did not assume all will go into full production; revenue guidance for the next two years is therefore discounted. The company also stated that if it secures more supply and client projects are deployed quickly, there’s still room to raise revenue outlooks.

Finally, regarding the AI networking business, management said that over the next few years, AI networking revenue is expected to maintain a similar growth rate to custom chips. Broadcom will also expand optical communication product capacity and proactively promote low-latency Ethernet technology. These network products will serve Broadcom’s custom chip clients and can also enter some GPU clusters, so the business will cover a broader AI infrastructure base.

Alright, that’s the core content. They also discussed client financing later, with similar intentions as Nvidia: now, for OA, for helping customers, as these firms grow, they will feed back into Broadcom’s business and support its long-term development, and so forth. No need to elaborate here.

Jason thinks the earnings call sent three important signals.

First, demand for AI semiconductors remains strong, but market expectations and share prices have run too far ahead, and time is needed for digestion.

Second, Google TPUs, whether in-house or via Anthropic, remain Broadcom’s largest contributor, but Google is already looking elsewhere (Marvell). Even if this is an incremental market and the pie is growing—everyone wins—investors will reprice for Broadcom’s permanently lost share and diminished bargaining power due to competition.

Third, expectations for the coming year are in line with the market, but those for the next two years are far apart, just like Nvidia’s guidance. What does this mean? It means the market has not fully priced in, or simply does not believe this cycle of AI semiconductors can last through 2028 or longer. The market is still pricing with inertia, using short-cycle thinking. There may not be a good solution; this will continue to suppress valuations for semiconductor firms, forcing a quarter-by-quarter adjustment to relieve the pressure.

Lastly, a thought on Anthropic. Broadcom has confirmed Anthropic’s in-house data center capabilities, with computing power reaching 10–15 GW over two years. From the perspective of a pure AI data center, that’s already a sizable scale.

From a competitive standpoint, if Anthropic could obtain US federal security certification or the trust of large enterprises, enabling it to directly serve enterprise clients, it would pose considerable pressure on Microsoft and Amazon. This may also explain why Microsoft and Amazon must embrace open source models and compete on cost-effectiveness and model variety to counter Anthropic.