In September 2026 Broadcom told investors its AI semiconductor revenue would roughly double to $115bn in fiscal 2027 and double again to $230bn in fiscal 2028. The stock closed at $349.57 on 28 September, 27.4% below its June high of $481.57 and up just 6.6% across twelve months. Set those two facts beside each other and the conventional reading — that the market doubts the guidance — is too lazy. The market is not disputing the revenue. It is repricing the quality of it. Broadcom’s trailing-twelve-month revenue across the entire company is $89.10bn. The fiscal 2028 AI number alone is 2.6 times that. A forecast of that shape does not get valued on a growth multiple, because it is not really a growth forecast — it is an order book, and an order book is worth exactly what its counterparties are worth. That is the distinction the 50 analysts covering this stock cannot agree on, and it is why their targets span $215.88 to $715.
A 3.3-times spread across a $1.67trn company is not normal, and it is the single most informative number on the sheet. For comparison, the equivalent spread on Microsoft is 2.0 times. What produces it is concentration. Broadcom’s AI revenue is not a market — it is a handful of named relationships: Google’s TPU volume, a 5-gigawatt TPU v8i commitment for Anthropic in 2027, and OpenAI’s Jalapeno accelerator. Shipyards have priced this problem for a century. A yard with a five-year order book from three shipping lines trades at a discount to a yard with the same revenue spread across forty customers, because the backlog is only as good as the balance sheets behind it, and a single counterparty rescheduling takes a visible bite out of a year. Broadcom is now a shipyard with a spectacular book and a very short customer list, and the equity market has started to price it accordingly.
Key facts
- AVGO closed at $349.57 on 28 September 2026 — up 6.6% over twelve months but 27.4% below its 52-week high of $481.57 set on 2 June 2026 — Nasdaq daily closes, retrieved 29 September 2026
- Q3 FY2026 AI semiconductor revenue $16.7bn, up 221% year on year and 54% sequentially; Q4 guided to $21.7bn, up 236% — CNBC, September 2026
- AI revenue outlook: roughly $58bn in FY2026, $115bn in FY2027, $230bn in FY2028 — Seeking Alpha, September 2026
- Company-wide TTM revenue $89.10bn (+48.7%), net income $38.27bn (+102.1%), market cap $1.67trn, P/E 44.6 — StockAnalysis, 29 September 2026
- Consensus target $531.85 across 50 analysts; street high $715, street low $215.88 — a 3.3x spread
- Morgan Stanley raised its target to $505 in early September 2026; Morningstar’s fair value estimate is $650
- Named AI counterparties include Google (Ironwood and TPU v8i), Anthropic (5GW of TPU v8i in 2027) and OpenAI (Jalapeno)
What is actually happening, and why the guidance is credible
Broadcom does not compete with Nvidia. It competes with the idea of buying from Nvidia. Its custom accelerator business — XPUs, in the company’s language — designs bespoke silicon for hyperscalers that would rather own their compute architecture than rent it at merchant-GPU margins. Google’s TPU line is the canonical example, and it has been running for a decade. What changed in 2026 is that the model went from one anchor customer to several, and the volumes went vertical.
The Q3 numbers are not ambiguous. “Demand for our custom AI accelerators and networking continues to be very strong. Q3 AI semiconductor revenue of $16.7 billion grew 221% year-over-year, and 54% quarter-over-quarter,” chief executive Hock Tan told investors. “In Q4 the momentum continues, and we expect AI semiconductor revenue to accelerate to $21.7 billion, up 236% year-over-year.” A 54% sequential increase at a $16.7bn quarterly base is not a demand signal that can be manufactured with channel stuffing.
Crucially, Broadcom has said the fiscal 2027 figure is backed by secured supply rather than pipeline. That is a materially stronger statement than a forecast. It means wafer capacity, advanced packaging and HBM allocation have been contracted, which in turn means someone has committed cash. Tan also indicated Broadcom would accelerate shipments of Google Ironwood TPUs destined for Anthropic and TPU v8i chips to Google, and would deliver tens of billions of dollars of processors to Google annually for several years. Google’s own compute ambitions have grown strange enough to include putting TPUs into orbit under Project Suncatcher, which is a useful reminder that the anchor customer is not running out of ideas about where to put silicon.
Supporting the whole structure is the memory and packaging supply chain, where the constraint economics have been brutal all year — the same dynamic that has driven Micron’s re-rating. Broadcom securing FY2027 supply in that environment is itself a competitive achievement.
The Anthropic wobble, and how the company answered it
On 14 September, CNBC reported that Broadcom’s chief executive had addressed a push by Anthropic to slow its buildout, and stated that the company’s AI revenue targets had not changed. That single exchange is the most important event in the stock’s September, and it explains the 5.2% monthly decline better than any macro story.
Consider what it revealed. A $230bn fiscal 2028 forecast rests on a small number of counterparties, and one of them was reported to be recalibrating its pace. Tan’s answer — targets unchanged — is exactly what a CEO with contracted supply and take-or-pay structures would say, and it may well be entirely accurate. But the market’s reaction was not about whether Tan was right. It was the realisation that a single customer’s scheduling decision is now a headline risk on a $1.67trn company. That is the concentration discount arriving in real time.
It is worth being fair to Anthropic here: a company that is simultaneously signing multi-gigawatt silicon commitments and committing $11.6bn to Akamai with a warrant struck at $111.33 is not retrenching. It is sequencing. But sequencing at the customer produces revenue timing risk at the supplier, and Broadcom’s guidance is annual, not cumulative.
The sceptics have found their voice too. Ed Zitron’s “The Haters Guide to Broadcom”, published on 20 September, circulated widely enough to reach the front pages of technical forums, and its argument is the durability argument in its sharpest form: that the AI backlog is a function of a small group of loss-making customers whose own funding is not yet proven. You do not have to accept the conclusion to notice that the bear case has finally acquired a coherent shape, which it lacked six months ago. The $215.88 street low is that argument, priced.
What the numbers imply at each end of the range
Broadcom has 4.77bn shares outstanding. That makes the arithmetic unusually clean.
At the $349.57 close, the market capitalisation is $1.67trn. Against the fiscal 2028 AI revenue guidance of $230bn, the stock currently trades at roughly 7.3 times AI revenue alone, before counting the $89.10bn of company-wide trailing revenue the AI business sits on top of. If you believe the $230bn, today’s price is not demanding. The trailing P/E of 44.6 sits on earnings that grew 102.1% year on year, which is also not, on its face, a bubble multiple.
At the street high of $715, the market cap becomes $3.41trn — Broadcom would need to roughly double from here, which a $230bn AI revenue year delivered on schedule would comfortably support on any conventional semiconductor multiple. At the street low of $215.88, the market cap falls to $1.03trn, destroying roughly $640bn of value and taking the trailing P/E to about 27. That is not a crash multiple. It is what Broadcom looks like valued as a very good, very large, cyclical semiconductor company rather than as an AI compounder — which is precisely the de-rating the bear case describes.
| The case for $715 | The case for $215.88 |
| FY2027 lands at or near $115bn on already-secured supply | One or more anchor customers reschedules and FY2027 comes in materially short |
| The new XPU customer signed in 2026 becomes a second Google-scale account | Customer count stays at three or four and the concentration discount persists |
| Custom silicon keeps taking share from merchant GPUs at the hyperscalers | Nvidia’s roadmap closes the total-cost gap and custom XPU economics narrow |
| Infrastructure software earnings cushion any semiconductor cyclicality | AI multiple compresses toward a normal semis multiple of roughly 27x trailing |
| Take-or-pay structures convert backlog into cash regardless of customer pace | Frontier-lab funding tightens and the loss-making customer base cannot pay |
Export controls are the unhedgeable variable
Every risk above is commercial and therefore modellable. The one that is not is policy.
Custom AI accelerators at the performance level Broadcom is shipping sit squarely inside the US export-control perimeter, and that perimeter has moved repeatedly and without much warning across the past three years. For Broadcom the exposure is subtler than it was for the merchant GPU vendors. Its customers are almost entirely American hyperscalers and American AI labs, so a direct China sales ban costs it comparatively little revenue. The exposure runs the other way: through the supply chain, and through where the compute physically lands.
Broadcom’s accelerators are fabricated in Taiwan and packaged across a distributed Asian supply chain. Any control regime that constrains advanced packaging capacity, HBM allocation or the movement of finished accelerators into particular jurisdictions hits delivery schedules, and delivery schedules are what the $115bn FY2027 number is made of. A customer building a 5-gigawatt cluster has to build it somewhere, and the list of jurisdictions where a US supplier can lawfully deliver that much accelerated compute has been getting shorter, not longer.
There is a second policy front that gets less attention: power. A 5-gigawatt commitment is a grid problem before it is a silicon problem. Interconnection queues in the major US datacentre markets now run to years, and local opposition to datacentre siting has become an electoral issue in several states. Broadcom can ship on time and still see revenue recognition slip because the customer’s substation is not energised. That risk is invisible in a semiconductor model and entirely real in the schedule.
What happens next
First, watch the customer count, not the revenue line. Broadcom added a new XPU customer in 2026 for the first time in roughly a year, and management has flagged additional prospects. Every incremental named account does more for the multiple than an equivalent amount of revenue from an existing one, because it directly attacks the concentration discount that produced the $215.88 street low. This is the single cheapest re-rating available to the company.
Second, expect the FY2027 number to be tested publicly at least once more before it is delivered. The Anthropic episode will not be the last of its kind, and Broadcom’s response template is now established: reaffirm the annual target, cite secured supply. That works until a quarter comes in below the implied run rate, at which point the market will conclude the take-or-pay protection is weaker than advertised. The first quarter that misses the sequential path to $115bn is the one that matters.
Third, on our reading the risk is skewed differently from the consensus. The $531.85 average target implies 52% upside and sits closer to the bull end of the distribution than the arithmetic justifies. We would frame it as: $715 requires FY2027 to land and the customer list to broaden — two conditions. $215.88 requires only that the market decides to value Broadcom as a semiconductor company rather than an AI company, which is a sentiment decision it can make in a fortnight and has already half-made, given the 27.4% drawdown from June. The guidance is probably right. Whether the market pays for it is a separate question, and it is the one that has actually determined the share price all year.
Frequently asked questions
What is the Broadcom stock prediction for 2026?
Our bull case is $715 and our bear case is $215.88, matching the street high and low. The consensus of 50 analysts is $531.85, implying 52.1% upside from the $349.57 close on 28 September 2026. The 3.3-times spread between the extremes is unusually wide for a company of this size.
Why did Broadcom stock fall if AI revenue is growing 221%?
Because the market has shifted from pricing the growth rate to pricing the durability of the revenue. Broadcom’s AI business depends on a small number of named counterparties, and reports in September that Anthropic was slowing its buildout showed how much a single customer’s schedule can move a $1.67trn company.
How much AI revenue does Broadcom expect?
Roughly $58bn in fiscal 2026, about $115bn in fiscal 2027 and approximately $230bn in fiscal 2028. Management has described the fiscal 2027 figure as underpinned by secured supply rather than pipeline. For context, company-wide trailing-twelve-month revenue is $89.10bn.
Who are Broadcom’s AI customers?
The publicly identified relationships include Google, for Ironwood and TPU v8i volume; Anthropic, with a 5-gigawatt TPU v8i commitment for 2027; and OpenAI, for its Jalapeno accelerator. Broadcom added a new XPU customer in 2026, its first in around a year.
Is Broadcom expensive at 44 times earnings?
On trailing earnings that grew 102.1% year on year, a 44.6 P/E is not obviously stretched. The stock trades at roughly 7.3 times its own fiscal 2028 AI revenue guidance. The bear case is not that the multiple is too high today but that AI revenue should carry a lower multiple than the market has been applying.
What would take Broadcom to $215.88?
A de-rating to roughly 27 times trailing earnings — the kind of multiple a large, high-quality cyclical semiconductor company carries. That implies a $1.03trn market cap. It does not require the AI revenue to disappear, only for the market to stop paying an AI premium for it.
This article is for information only and is not investment advice. Price targets are scenario analysis, not forecasts.