Key points
- Broadcom (AVGO) reports fiscal Q3 on Wednesday, Sept. 2, with AI chip revenue guided to $16 billion, up over 200%.
- The stock closed Friday at $368.68, 25% below its June high, after its last record quarter sent shares down 12.6% in a day.
- The options market prices about a $30 move, just over 8%, through the Friday after the report.
Broadcom (AVGO) will release fiscal third-quarter results after the closing bell on Wednesday, Sept. 2, and the bar is already set. The company guided revenue to about $29.4 billion, an 84% increase from the prior year. It expects AI chips to contribute $16.0 billion, more than three times the year-earlier amount. Broadcom enters the report at 25% below its June high. That discount reflects what happened after the previous report: even with record results, the shares dropped 12.6% the following day.
What did Broadcom promise for the quarter?
The guidance came with the second-quarter report on June 3. That quarter, ended May 3, brought in $22.2 billion in revenue, up 48%, and adjusted earnings of $2.44 a share, up 54%. "Q2 semiconductor revenue from AI of $10.8 billion grew 143% year-over-year, above our forecast, driven by increasing demand for custom AI accelerators and AI networking," chief executive Hock Tan said in the release.
| Metric | Q3 FY2026 guidance | Change from a year ago |
|---|
| Revenue | About $29.4 billion | Up 84% |
| AI semiconductor revenue | $16.0 billion | Up over 200% |
| Adjusted operating margin | 67% of revenue | |
Wall Street is looking for adjusted earnings of about $3.16 per share, compared with $1.69 in the same quarter a year ago. During the June earnings call, Broadcom also projected that fiscal 2027 AI revenue would top $100 billion. Its target for this year is $56 billion.
Where does the OpenAI deal show up?
OpenAI and Broadcom disclosed their partnership in October 2025. The plan calls for 10 gigawatts of custom accelerators designed by OpenAI, with rack deployments scheduled to begin in the second half of 2026 and conclude by the end of 2029. This report covers the first quarter of that window.
The partnership's first chip has already been named and benchmarked. On June 24, OpenAI introduced Jalapeño, an inference processor developed with Broadcom. At the Hot Chips conference on Aug. 25, OpenAI published benchmarks it says show the chip doing more AI work per unit of power than Nvidia hardware. SemiAnalysis, which watched the runs in person, put Jalapeño's output-token throughput per megawatt above Nvidia's next-generation Vera Rubin, while calling the two a dead heat on cost per token. Broadcom did the physical chip design and supplies the Tomahawk networking that connects the racks. TSMC makes the chip; Celestica assembles the systems.
For Broadcom, the design behind the efficiency claim is the key detail. OpenAI explained at Hot Chips that Jalapeño uses 64 independent core slices, each with dedicated memory, rather than the single shared memory pool used by a GPU. GPUs hide memory delays by running thousands of threads simultaneously. That approach performs well on large batches but loses its advantage when the system is responding to one user. Serving a model is mostly that second case. At a concurrency of one, OpenAI measured more than 700 tokens per second per user on DeepSeek R1 and roughly 1,400 on Kimi K2.5.
Another part of OpenAI's presentation maps directly onto Broadcom's revenue categories. The chip design includes a dedicated network for collective communication, separate from the general-purpose interconnect, while Broadcom's Tomahawk 6 switches connect the racks. In this system, networking is part of the core architecture, not an add-on. Hock Tan divides AI revenue along the same lines: custom accelerators and networking. An OpenAI deployment can therefore contribute to Broadcom's results through both categories. OpenAI also reported that Jalapeño moved from RTL to tapeout in nine months, with Broadcom completing the physical design.
Financing activity around Broadcom's AI operation accelerated throughout the summer. The company is reportedly seeking more than $60 billion in financing connected to one AI transaction. In July, Samsung signed a $200 billion foundry agreement with Broadcom. After a month of debate over AI capital spending, the market is likely to focus closely on anything Tan says about the pace of OpenAI rack shipments.
Why is the stock 25% below its June high?
Broadcom reached $495 on June 3, then released its record second-quarter results that afternoon. In the next trading session, the stock sank 12.6%, closing at $418.91 after starting from $479.23. The decline continued over the summer, leaving the shares at $368.68 on Friday.
Nvidia (NVDA) may have improved the tone this week. It reported $96.2 billion in revenue on Aug. 26, up 106%, and its shares jumped 8.7% the next day, the biggest one-day gain since April 2025. Broadcom advanced 4.5% in the same session, then surrendered part of that increase on Friday.
The options market has its number for Wednesday. A $370 straddle expiring Sept. 4, the trade that pays if the stock moves far in either direction, cost about $30 at Friday's close, a move of just over 8%. The earnings call starts at 2 p.m. Pacific, 5 p.m. Eastern.
