Author: Rick Orford
Compiled by Deep潮 TechFlow
DeepInsight Summary: Broadcom has secured approximately $350 billion in AI chip shipments over the next two fiscal years; top clients such as Anthropic and OpenAI, while publicly supporting the notion of slowing AI development, continue to place custom chip orders at gigawatt-scale. Rick Orford believes the demand narrative remains intact, with valuation multiples having significantly declined alongside improved profitability; he maintains a Strong Buy rating while highlighting financing guarantees and memory costs as key risks.
Broadcom Inc. (AVGO) informed investors that it expects to ship approximately $350 billion worth of AI chips over the next two fiscal years, with supply already secured. This visible and well-defined demand underscores just how strong the company’s AI business is right now.
However, a striking opposing voice has just emerged: Anthropic (ANTHRO) CEO Dario Amodei recently proposed slowing down AI development for public safety—surprising nearly everyone, other major figures in the AI industry, including Sam Altman of OpenAI (OPENAI) and Elon Musk of SpaceX (SPCX), almost immediately agreed.
Nevertheless, I still don't believe this poses an immediate, substantial threat to Broadcom's AI business. The warnings aim to control the pace of AI advancement and manage risk, not to halt it; meanwhile, demand for data centers to run and scale existing models continues to far exceed supply.
Therefore, I believe the narrative around demand for Broadcom's chips remains valid, and I maintain a strong buy rating. The reasons are as follows.
In my previous analysis, I argued that the market misinterpreted a "reaffirmation" as a "ceiling." Since then, management raised its FY2027 AI chip revenue target by 15%, introduced a new FY2028 target of $230 billion that did not exist in June, and increased this fiscal year’s AI revenue forecast to $58 billion.
What’s changing now is that Broadcom is measuring demand in gigawatts rather than chip orders. This makes sense—what’s truly limiting customers is how much power their data centers can draw.
As confirmed during the Q3 earnings call, Anthropic will deploy 1 gigawatt of Ironwood chips this year, 5 gigawatts of the next-generation TPU in 2027, and an additional 10 gigawatts in 2028—making Anthropic Broadcom’s largest custom chip customer next year.
Meanwhile, OpenAI has reserved 1.3 gigawatts of Jalapeño chips for 2027, exceeding 5 gigawatts the following year. Meta (META) will consume three generations of MTIA chips before 2027, and Alphabet (GOOG/GOOGL) has signed a long-term agreement—Broadcom CEO Hock Tan described the scale as “hundreds of billions of dollars annually”—which is very bullish, but I’d prefer to hear specific figures.
Tan stated that each gigawatt of computing power generates $20 to $30 billion in "content" revenue for Broadcom, and expects this ratio to remain relatively stable. Here, "content" refers to the dollar value of Broadcom’s own products—primarily custom AI accelerators and networking chips—that customers purchase for every gigawatt of AI data center capacity they build.
Bernstein’s Stacy Rasgon noted that the targets of $11.5 billion and $23 billion, in relation to the approximately 30 gigawatts of demand mentioned in the call, equate to roughly $11 billion or $12 billion per gigawatt. Tan’s response was straightforward: not all 30 gigawatts will come online within those two fiscal years, as land, construction, and power infrastructure require time.
Let’s do a rough calculation. Tan expects total shipments of $350 billion in AI chips for FY2027 and FY2028. Dividing this by his own “content” figures, it translates to approximately 12 to 17 gigawatts of actual deployment, while customers are aiming for 30 gigawatts.
This means that before presenting the numbers, management had already cut the roadmap roughly in half. This is the opposite of the typical approach to multi-year chip forecasts, which usually starts by stating the maximum number the market could theoretically absorb, then assumes most of that will become sales.
When a company sets aside such a large buffer in advance, this forecast reads more like a "floor": you only miss it if deployment is slower than what management has already assumed.
The order book I wrote about in July has begun to materialize: from $10.8 billion shipped versus $30 billion booked, to $16.7 billion, and now guidance of $21.7 billion—all confirmed in the same earnings call.
Management also stated that Broadcom began shipping Alphabet’s 8th-generation TPU before MediaTek’s v8t, despite MediaTek starting earlier. The competitive risks I previously highlighted have not disappeared, but in this generation of chip competition, Broadcom is the clear winner.
At the time, I viewed Broadcom’s fabless model as a lasting advantage, evidenced by quarterly capital expenditures of approximately $230 million. This still holds true, but the narrative requires refinement: Q3 capital expenditures reached $532 million, roughly a 1.3-fold increase quarter-over-quarter, with guidance for Q4 at $1.4 billion—about a sixfold increase over two quarters.
Where did the money go? Management confirmed on the earnings call that Broadcom is building a substrate factory in Singapore—substrates being the base materials chips are mounted on—and is increasing production capacity for optical lasers used for inter-rack data transmission by more than threefold; these lasers are essential once copper wiring reaches its limits.
Tan listed these two items alongside memory chips as anticipated bottlenecks, so the company is effectively "bringing back" part of its own manufacturing process—taking matters into its own hands where money alone cannot solve the problem. This is a reasonable step toward achieving its revenue targets.
The scale remains manageable: the $1.4 billion in capital expenditures outlined is approximately 4% of the $34.8 billion quarterly revenue guidance. Unable to resolve its memory needs internally, Broadcom signed a memory and manufacturing agreement with Samsung in July, worth over $200 billion and extending through 2030. Currently, this is only a memorandum of understanding, so I am not incorporating it into my model; however, it clearly signals that Broadcom is securing coverage for the one critical input it cannot produce itself.
I previously noted that the gross margin declined to 77.1% and continued to fall. The third quarter reported 75%, with guidance for the fourth quarter nearing 73%—compared to 78% in the same period last year—due entirely to higher memory costs. Custom AI chips require more memory, and their share in the sales mix continues to rise: AI now accounts for 56% of total revenue, up from 49% in the second quarter.
According to reports, server memory contract prices have risen by as much as 270% this year. The significant decline in gross margin is largely due to memory shortages being passed through to Broadcom’s cost line, rather than the company losing pricing power.
An even more impressive aspect is that the company’s operating profit margin reached a record 67.9%, with guidance for the fourth quarter at 66%, flat compared to last year. Looking specifically at the chip business, the operating profit margin stood at 61%, driven by revenue growth of 127%, far outpacing operating expenses, which rose only 22%. When analysts asked about future gross margins, Tan told the audience to stop focusing on gross margin and instead look at operating profit margin—because revenue growth has consistently outpaced the growth in operating costs.
At first, I thought it was convenient for the CEO to say this amid declining gross margins, but the numbers hold up—I’ll give him that.
So how much is all of this worth? Let’s do the math. Revenue for the first three quarters has already reached $71.1 billion, plus the Q4 guidance of $34.8 billion, bringing FY2026 close to $106 billion—compared to last year’s revenue of $63.89 billion, this implies an implied growth rate of approximately 66%.
Profits accumulate along the same path. Broadcom has already recorded $7.81 per share in the first three quarters, and the fourth quarter appears to add approximately another $3.80—this estimate is derived by applying the guided 66% operating margin to the $34.8 billion fourth-quarter revenue guidance, then subtracting interest and taxes.
This results in an adjusted profit of approximately $11.60 for FY2026. At the current stock price of $344.72, the stock is trading at approximately 31 times forward P/E (non-GAAP).
If the same algorithm is extended to FY2027—when management states that supply is already locked in—AI would contribute $115 billion in guidance, non-AI chip business approximately $17.5 billion, and software business approximately $35 billion, totaling roughly $168 billion in revenue.
If the operating margin remains at 66% and the tax rate is 16%, Broadcom could earn approximately $18 per share; if the operating margin declines to 62% (memory costs could easily hurt more than currently expected), it would be around $17. At today’s stock price, this corresponds to about 19 times earnings for the $18 scenario and slightly over 20 times for the $17 scenario; all these assumptions are directly derived from the Q3 earnings call.
Tan went further, stating that Broadcom is on track to achieve its target of over $30 per share by FY2028, which equates to approximately 11.5 times today’s stock price—this is a goal set by management itself. In July, I was still writing about a forward multiple of around 42, compared to a five-year average of about 48; now, the multiple has clearly declined, while the underlying earnings base has risen.
I believe the market is no longer debating whether demand exists. It is now questioning whether Broadcom can convert that demand into actual payments—this issue will be addressed in the next section.
This is only the optimistic half. Now let’s look at the less optimistic side. In June, Broadcom partnered with Apollo (APO) and Blackstone (BX) to build the AI XPV platform, aiming to raise over 20 gigawatts of computing power for AI labs by 2028; the first tranche, closed that same month, amounted to $35 billion and covers the previously mentioned Anthropic 1-gigawatt deployment.
Anthropic has now reached 16 gigawatts, yet no one has clarified how much more funding is needed to complete the journey. The mechanism is worth understanding: large investors purchase entire racks of chips and lease them to labs, so Broadcom lends not a single dollar. The company provides a guarantee for rental payments—and it is precisely this guarantee that enables the safest portion of the debt to receive higher ratings and lower financing costs.
Documents show that Broadcom could lose up to $29 billion in the worst-case scenario of the first round—if the lab completely stops payments and the equipment ultimately becomes worthless. When analysts asked whether this would be the ongoing cost per gigawatt, the CFO declined to provide an upper limit.
In my view, the $29 billion figure is “relatively soft” because it relies on the assumption that only if devices retain residual value does this number hold up—yet Broadcom’s own product roadmap is precisely undermining the resale value it just guaranteed.
Ironwood TPU just shipped this season, and the next generation of Google’s chips is already in production—offering a five-year warranty on hardware value while sitting next to a company that refreshes hardware annually; this feels somewhat incongruous. The secondary market for custom accelerators is at best thin. Broadcom is essentially insuring the resale value of its chips while personally undermining that value with each new design iteration.
Management’s response was: “Anthropic is moving toward an IPO, and credit will improve”—that’s all they said. An IPO may reduce the probability of customer default, but it does nothing to mitigate the extent of losses after a default occurs, which is precisely what the equipment residual value assumption addresses.
What I’m concerned about isn’t customer default, but rather that cheap money is cheap only because Broadcom continues to stand behind it with guarantees; refusing to sign a guarantee could raise financing costs and slow the deployment pace needed to support its $115 billion AI revenue target, with little buffer elsewhere.
Despite the CEO and department heads jointly calling for a brake on AI development, demand for Broadcom’s products should remain intact. Once again: the industry agrees to set speed bumps, not to completely halt AI development. Therefore, the bullish narrative still holds.
The final decision rests on price. Investors are assuming credit risk at the business's edge while paying multiples lower than in July, yet corresponding to a larger profit base. Taking all this into account, I maintain my strong buy rating on Broadcom.