Broadcom Sees Second AI Demand Wave Reaching Enterprises via APIs

Yes—management expects a “second wave” as AI reaches enterprises/cloud, but it largely flows through the same frontier labs (CaaS/APIs) that drive hyperscaler demand

1) What would the “second wave” look like?

Management acknowledges the timing is shifting: initially, most AI compute was tied to hyperscalers running their own customer workloads, but the call suggests a new layer is forming as enterprises consume AI via tokens and cloud services/APIs become widely usable. The participant framing explicitly asks whether AI hitting enterprises and consumers creates a “big second wave of demand” beyond hyperscalers. ¹

Broadcom’s response reframes the demand source rather than denying growth:

• Enterprise demand for AI is still early, but it is beginning to show up. ²
• Enterprises consume AI via tokens pulled from the same large platform/API providers—specifically naming Anthropic, OpenAI, and Gemini—and management believes “substantially most” token consumption is tied to those LLMs. ²
• As frontier model labs “productize” their models (e.g., Opera/ChatGPT/Gemini examples cited in the transcript), the end demand on compute capacity ultimately converges on the same compute capacity Broadcom supplies to those few labs. ²

So the “second wave” is expected in who is consuming the applications (enterprises/consumers) more than in who is originating the compute demand (the frontier labs). ²³

2) How does Broadcom compare enterprise/cloud-driven demand to hyperscaler demand?

Broadcom’s position is that cloud delivery does not fundamentally change the compute demand chain:

• Management says that for AI delivered as SaaS/APIs (examples include Bedrock, Vertex, Azure and first-party clouds), “most of the demand at the end of the day… comes from those few large frontier model developers.” ³
• They emphasize that demand is unlikely to shift into enterprises buying/operating their own XPUs at scale; most enterprises are not building their own full stack and running compute themselves, so compute still routes back to the frontier-lab ecosystem. ³
• A key supporting point is that they observe “insatiable growth” in compute capacity as enterprises start using AI tokens and as consumers buy from the same source; they expect this demand to continue through 2027 and 2028. ²

Conclusion on the comparison: hyperscalers may have led the early wave for their own workloads, but the incremental enterprise/cloud wave is described as largely amplifying the same frontier-lab compute pulls, not replacing them. ²³

3) Gigawatt commitments: enterprise/cloud wave vs hyperscaler wave (what the excerpt supports)

The transcript distinguishes “two classes of customers” in the context of gigawatt commitments—not primarily hyperscalers vs enterprises, but rather:

• Two CaaS-like providers (Anthropic/OpenAI) are said to have “very large gigawatt commitments in the out years,” while
• the “oldest customer” (and “Google’s” offerings are mentioned) is smaller in the comparison, though with catch-up dynamics for later entrants. ¹

Although the excerpt you provided does not provide a full numeric table for every customer class, it does include concrete gigawatt commitment figures for specific frontier-lab customers and thus anchors the “gigawatts” discussion:

Named frontier-lab commitments (examples directly stated):

• Anthropic: “access to Broadcom TPU based compute of over 1 gigawatt” for 2026, plus “another 5 gigawatts … beginning in 2027.” ⁴
• OpenAI: contractual commitment to “deploy 1.3 gigawatts in 2027” as part of a larger “10 gigawatts by 2029” agreement. ⁴
• Meta: expects “to deploy 3 gigawatts through the end of 2028,” with “Initial order for 1 gigawatt … will start delivery in the second half of 2027.” ⁴

Broader “visibility” and timing support for gigawatt ramp persistence:

• Management describes lead times and planning needs (including power) and states visibility now runs “all the way to 2028,” up from “pretty much … to 2027” three months earlier. ⁵
• They also state compute capacity demand is expected to continue accelerating for certain customers and that “for even 2 of our customers, … Anthropic and OpenAI,” compute power measured in gigawatts is already “way ahead” of what they expected six months earlier; they also say gigawatts in totality (including other customers) will keep growing, with 2028 “a substantial growth from what we are forecasting in 2027.” ⁶

Bottom line on gigawatts (enterprise/cloud wave vs hyperscaler demand):

• The excerpt supports that the largest out-year gigawatt commitments belong to frontier labs that deliver via consumer/enterprise-accessible APIs and clouds (Anthropic/OpenAI explicitly called out as having “very large gigawatt commitments”). ¹
• Enterprise and consumer adoption is described as drawing tokens from these same frontier labs, so the enterprise/cloud “second wave” is expected to translate into additional compute gigawatt requirements largely aligned with those frontier-lab commitments, not a separate set of gigawatts driven by enterprises directly buying hardware. ²³
• Therefore, in terms of gigawatt commitments, the excerpt suggests the “dominant” gigawatt pipeline is still frontier-lab-driven (which may be delivered to end customers through hyperscalers/cloud), rather than enterprise/local demand creating an entirely independent gigawatt order book. ¹²³

Overall answer

• Expectation of a second wave: Yes, management expects demand to broaden as AI consumption reaches enterprises and consumers, but they describe it as being powered by token consumption from the same few frontier model providers (Anthropic/OpenAI/Gemini). ²³
• Comparison to hyperscaler demand in gigawatts: The excerpt indicates that the largest out-year gigawatt commitments are with frontier labs offering “consumer versions” of AI (Anthropic/OpenAI), and management believes enterprise/cloud usage routes back to that compute demand chain. ¹²³
• Gigawatt commitment examples: Broadcom cites specific frontier-lab compute deployment commitments of >1 GW (Anthropic 2026) and another 5 GW (Anthropic 2027), 1.3 GW (OpenAI 2027) with a larger 10 GW by 2029, and 3 GW through end of 2028 (Meta). ⁴

This framework implies that while hyperscalers may have dominated the early cycle, the “second wave” is best viewed as incremental amplification of gigawatt needs from the frontier-lab ecosystem rather than a wholly different gigawatt commitment profile sourced from enterprises directly. ²³