Broadcom CEO Holds $350B Outlook Strong Amid Calls to Slow Froniter AI Development

NanoViricides Advances Ebola Drug Toward Phase II Trial in DRC (From Equiscreen)

Concerns around security and how artificial intelligence capabilities could evolve have put a spotlight on the pace of frontier model development. After doing so several months ago, Anthropic has again called for measures that would limit how quickly frontier models can improve.

This has understandably caused significant concerns about the AI trade. On Sept. 14, the next trading day after Anthropic CEO Dario Amodei released his blog post “We Must Pace the Frontier," many AI-related stocks tanked.

Broadcom (NASDAQ: AVGO) fell 4.8%, and higher volatility names like Astera Labs (NASDAQ: ALAB) fell more than 10%. Markets expect these stocks to put up drastic sales and earnings growth going forward, something that a slowdown in frontier model development could threaten.

Broadcom CEO Hock Tan recently faced direct questions about how pacing AI development could affect its approximately $350 billion AI sales outlook. Unsurprisingly, Tan remained firm on his forecast, with his comments revealing a real nuance in the AI-slowdown debate—a nuance that happens to favor Broadcom.

Broadcom Brushes Off AI Development Slowdown Concerns

In its last earnings call, Broadcom guided for $115 billion of AI chip revenue in its fiscal year 2027 (FY2027), and $230 billion in fiscal year 2028 (FY2028). Rounding up a few billion dollars, Hock Tan said “we believe with a pretty high degree of confidence, we will ship $350 billion of AI semiconductors to these customers in the next 2 years.”

When asked "Is there anything that has happened in this whole debate about the AI slowdown that would give you pause to that prediction?" Tan said, “No, not in the least." Tan added that the company sees compute demand for frontier model development (training) and inference as continuing to be very strong and durable. Training is the process of making models more intelligent, whereas inference refers to using already trained models to execute tasks.

However, Tan made a slightly more revealing statement later on, saying “Look, I don’t know about training, but when you want to productize inference, I see it continuing to be very, very strong." This statement lines up with how an agreed-upon slowdown in AI development would likely affect the industry.

Models becoming too intelligent, too quickly, is the primary fear discussed by Anthropic and others. In turn, a slowdown in AI development would likely have the most negative effect on demand for training compute. In fact, Amodei specifically notes in his blog, “We should also consider pacing based on limiting the ingredients that go into frontier models, such as training compute." Luckily for Broadcom, its AI chip business is becoming more inference oriented than training oriented, demonstrated by its customer relationships.

Broadcom’s Top Customer Relationships Focus on Inference

In its latest earnings call, Broadcom noted that it is Alphabet’s (NASDAQ: GOOGL) partner in developing the firm’s next-generation tensor processing unit (TPU) v8i. This is the inference-optimized variant of the TPU v8, while MediaTek (OTCMKTS: MDTKF) is Alphabet’s partner in developing the training-optimized TPU v8t. Thus, Broadcom is clearly more exposed to inference demand with this chip.

The case is the same when it comes to the firm’s collaboration with OpenAI. Broadcom has helped OpenAI develop Jalapeño, which the firms explicitly call “OpenAI’s first custom inference chip.” While revenue from the OpenAI relationship is likely limited at this point, Broadcom does not expect this to be the case for long. As Jalapeño rolls out, Broadcom expects OpenAI to become the company’s second-largest custom chip buyer in FY2028.

Then there is Anthropic, which Broadcom expects to become its largest customer chip revenue source in FY2027 and sustain this position in FY2028. Broadcom is not developing a separate chip for Anthropic; the company will also deploy TPUs. Nonetheless, these chips will also be inference-optimized, with Anthropic expected to deploy 5 GW of the TPU v8i in 2027 and 10 more GW in 2028.

Calls for AI Slowdown Remain a Risk to Watch

There is reason to believe that training compute demand would be most adversely affected if the pace of frontier model development were to slow. In this case, Broadcom’s heavy focus on inference compute could leave it in a better position.

None of this is to say that a slowdown in AI development would not negatively affect Broadcom's business and stock price. It very well could, especially if slowing down development causes frontier models to lose their intelligence lead, thereby shifting their share of inference demand. In turn, it is important to continue monitoring whether calls for a slowdown in frontier development intensify and move toward implementation.

Read this article online ›