$115B for FY27, $230B for FY28, $30 of EPS: Broadcom Doubles the Framework Twice and the Stock Falls on a 0.7% Guide Shortfall
AVGO financial model
Income Statement · Dollars in millions, except per share
| Income Statement | ||||||
|---|---|---|---|---|---|---|
| Actual | Estimate | |||||
| FY23 | FY24 | FY25 | FY26E | FY27E | FY28E | |
| Net Revenue — Semiconductor Solutions | $28,182.0 | $30,096.0 | $36,858.0 | $75,911.9 | $128,812.3 | $175,184.7 |
| Net Revenue — Infrastructure Software (VMware from Q1'24) | 7,637.0 | 21,478.0 | 27,029.0 | 29,282.6 | 32,201.1 | 35,099.2 |
| Total Net Revenue | $35,819.0 | $51,574.0 | $63,887.0 | $105,194.5 | $161,013.4 | $210,283.9 |
| Check: Products + Subs − Total (must = 0) | 0.000 | 0.000 | 0.000 | |||
| Less: Cost of Products Sold | ($8,636.0) | ($9,797.0) | ($12,115.0) | ($22,503.8) | ($35,947.3) | ($48,365.3) |
| Less: Cost of Subscriptions & Services | (636.0) | (2,991.0) | (2,371.0) | (3,057.5) | (4,258.0) | (5,257.1) |
| Less: Amortization of Acquisition-Related Intangibles (COGS) | (1,853.0) | (6,023.0) | (6,031.0) | (5,779.6) | (5,460.5) | (4,914.5) |
| Less: Restructuring Charges (COGS) | (4.0) | (254.0) | (76.0) | (23.0) | 0.0 | 0.0 |
| Total Cost of Revenue | ($11,129.0) | ($19,065.0) | ($20,593.0) | ($31,363.9) | ($45,665.8) | ($58,536.9) |
| Gross Margin | $24,690.0 | $32,509.0 | $43,294.0 | $73,830.6 | $115,347.6 | $151,747.0 |
| Less: Research & Development | ($5,253.0) | ($9,310.0) | ($10,977.0) | ($12,794.9) | ($15,688.0) | ($18,505.0) |
| Less: Selling, General & Administrative | (1,592.0) | (4,959.0) | (4,211.0) | (4,807.9) | (6,284.4) | (7,359.9) |
| Less: Amortization of Acquisition-Related Intangibles (OpEx) | (1,394.0) | (3,244.0) | (2,031.0) | (2,017.4) | (1,978.9) | (1,879.9) |
| Less: Restructuring & Other Charges | (244.0) | (1,533.0) | (591.0) | (174.0) | 0.0 | 0.0 |
| Total Operating Expenses | ($8,483.0) | ($19,046.0) | ($17,810.0) | ($19,794.2) | ($23,951.3) | ($27,744.9) |
| Operating Income | $16,207.0 | $13,463.0 | $25,484.0 | $54,036.4 | $91,396.3 | $124,002.2 |
| Less: Interest Expense | ($1,622.0) | ($3,953.0) | ($3,210.0) | ($3,137.0) | ($3,120.0) | ($3,120.0) |
| Other Income (Expense), Net | 512.0 | 406.0 | 455.0 | 791.0 | 480.0 | 480.0 |
| Income Before Income Taxes | $15,097.0 | $9,916.0 | $22,729.0 | $51,690.4 | $88,756.3 | $121,362.2 |
| Less: Provision for (Benefit from) Income Taxes | ($1,015.0) | ($3,748.0) | $397.0 | ($5,336.2) | ($10,650.8) | ($15,777.1) |
| Income from Continuing Operations | $14,082.0 | $6,168.0 | $23,126.0 | $46,354.2 | $78,105.6 | $105,585.1 |
| Income (Loss) from Discontinued Operations, Net | 0.0 | (273.0) | 0.0 | 0.0 | 0.0 | 0.0 |
| Net Income | $14,082.0 | $5,895.0 | $23,126.0 | $46,354.2 | $78,105.6 | $105,585.1 |
| Less: Dividends on Preferred Stock | $0.0 | $0.0 | $0.0 | $0.0 | $0.0 | $0.0 |
| Net Income Attributable to Common Stock | $14,082.0 | $5,895.0 | $23,126.0 | $46,354.2 | $78,105.6 | $105,585.1 |
| EPS — Basic | $3.39 | $1.27 | $4.91 | $9.76 | $16.34 | $21.96 |
| EPS — Diluted | $3.30 | $1.23 | $4.77 | $9.49 | $15.91 | $21.38 |
| Shares — Basic | 4,150.0 | 4,624.0 | 4,712.0 | 4,750.8 | 4,779.1 | 4,807.8 |
| Shares — Diluted | 4,270.0 | 4,778.0 | 4,853.0 | 4,884.5 | 4,909.0 | 4,938.5 |
| Ratios & Assumptions | ||||||
| YoY Total Net Revenue Growth | 7.9% | 44.0% | 23.9% | 64.7% | 53.1% | 30.6% |
| Gross Margin % | 68.9% | 63.0% | 67.8% | 70.2% | 71.6% | 72.2% |
| Operating Margin | 45.2% | 26.1% | 39.9% | 51.4% | 56.8% | 59.0% |
| Net Margin | 39.3% | 11.4% | 36.2% | 44.1% | 48.5% | 50.2% |
| Effective Tax Rate | 6.7% | 37.8% | (1.7%) | 10.3% | 12.0% | 13.0% |
The full workbook adds 22 historical and 6 projected quarters, plus KPI Drivers · Balance Sheet · Cash Flow Statement — every subtotal a live formula, every projection traced to a driver.
Key Takeaways
- Every line beat. Revenue of $29.59B (+86% YoY) topped both the company's own ~$29.4B guide and the $29.36B to $29.43B consensus range; AI semiconductor revenue of $16.7B (+221% YoY, +54% QoQ) cleared management's $16B guide by $0.7B and the Street's ~$15.2B estimate by roughly 10%; non-GAAP EPS of $3.32 beat the $3.24 consensus. Free cash flow was a record $13.67B at 46% of revenue.
- The out-year framework moved a long way. FY26 AI revenue was raised to $58B from ~$56B, FY27 went from "in excess of $100B" to approximately $115B, and FY28 was introduced at approximately $230B, all described as supply-secured. Management also put its first explicit out-year earnings marker on the record: on target to exceed $30 of EPS in fiscal 2028.
- The customer hierarchy inverted. Anthropic is now expected to become the largest XPU customer in 2027 and hold that position in 2028 (5 GW of TPU v8i in 2027, a further 10 GW of line-of-sight in 2028), with OpenAI second on more than 5 GW in 2028. Google, the anchor for a decade, was reframed as "multi-tens of billions of dollars of TPUs annually," which is enormous but no longer the largest single line.
- The one thing that got worse is the financing question. An analyst put the last 10-Q's ~$29B maximum exposure on the first XPV tranche on the record and management confirmed those numbers "remain true," then declined to give any aggregate cap for a platform designed to fund more than 20 GW. That is roughly 29% of book equity attached to a single gigawatt, with no disclosed ceiling.
- Rating: Maintaining Outperform. At the $367.24 pre-print close the stock trades near 32x our FY26 non-GAAP EPS estimate and roughly 12x management's own FY28 marker, having de-rated 23% since the FQ2 print while the contracted framework expanded twice. We continue to underwrite this against the supply-secured FY27 number rather than the FY28 number, and we now treat the undisclosed XPV exposure as the thesis's live risk rather than a watch item.
Results vs. Consensus
FQ3 FY2026 Scorecard
| Metric | Actual | Consensus / Guide | Beat/Miss | Magnitude |
|---|---|---|---|---|
| Revenue | $29.591B | $29.36B–$29.43B cons.; ~$29.4B guide | Beat | +0.6% to +0.8% |
| AI semiconductor revenue | $16.7B | ~$15.2B cons.; $16.0B guide | Beat | +9.9% vs. cons.; +4.4% vs. guide |
| Non-GAAP gross margin | 75.0% | ~74% guide | Beat | +100bps vs. guide |
| Non-GAAP operating margin | 67.9% | ~67% guide | Beat | +90bps vs. guide |
| Non-GAAP operating income | $20.095B | n/a | Record | +92% YoY |
| EPS (non-GAAP, diluted) | $3.32 | $3.22–$3.24 | Beat | +$0.08 to +$0.10 |
| EPS (GAAP, diluted) | $2.68 | ~$2.55 | Beat | +215% YoY |
| Free cash flow | $13.665B | n/a | Record | 46.2% of revenue |
| Non-AI semiconductor revenue | ~$4.2B | ~$4.5B guide | Miss | +5% YoY vs. +12% guided |
| Infrastructure software revenue | $8.752B | ~$8.9B guide | Slight miss | +29% YoY vs. +31% guided |
| FQ4 revenue guide | ~$34.8B | $35.03B–$35.1B cons. | Miss | -0.7% to -0.9% |
Year-Over-Year Comparison
| Metric ($M except per-share) | FQ3 FY26 | FQ3 FY25 | YoY Change |
|---|---|---|---|
| Net revenue | 29,591 | 15,952 | +86% |
| Semiconductor solutions | 20,839 | 9,166 | +127% |
| Infrastructure software | 8,752 | 6,786 | +29% |
| Non-GAAP gross margin | 22,191 (75.0%) | 12,499 (78.4%) | +78%; -336bps margin |
| Non-GAAP operating expenses | 2,096 | 2,044 | +2.5% |
| Non-GAAP operating income | 20,095 (67.9%) | 10,455 (65.5%) | +92%; +237bps margin |
| GAAP operating income | 15,955 | 5,887 | +171% |
| Non-GAAP net income | 16,372 | 8,404 | +95% |
| GAAP net income | 13,088 | 4,140 | +216% |
| Non-GAAP diluted EPS | $3.32 | $1.69 | +96% |
| GAAP diluted EPS | $2.68 | $0.85 | +215% |
| Stock-based compensation | 2,019 (6.8% of rev) | 2,322 (14.6% of rev) | -13% |
| Cash from operations | 14,197 | 7,166 | +98% |
| Free cash flow | 13,665 (46.2%) | 7,024 (44.0%) | +95%; +215bps margin |
Quarter-Over-Quarter Comparison
| Metric ($M except per-share) | FQ3 FY26 | FQ2 FY26 | QoQ Change |
|---|---|---|---|
| Net revenue | 29,591 | 22,187 | +33.4% |
| AI semiconductor revenue | ~16,700 | ~10,800 | +54% |
| AI as % of total revenue | 56% | 49% | +7pts |
| Non-GAAP gross margin | 22,191 (75.0%) | 17,109 (77.1%) | +29.7%; -212bps margin |
| Non-GAAP operating expenses | 2,096 | 2,181 | -3.9% |
| Non-GAAP operating income | 20,095 (67.9%) | 14,928 (67.3%) | +34.6%; +63bps margin |
| Non-GAAP net income | 16,372 | 12,074 | +35.6% |
| Non-GAAP diluted EPS | $3.32 | $2.44 | +$0.88 (+36%) |
| Free cash flow | 13,665 | 10,262 | +33.2% |
| Capital expenditures | 532 | 231 | +130% |
| Cash and equivalents | 23,975 | 19,628 | +4,347 |
| Inventory | 4,523 | 4,328 | +4.5% |
FQ2 inventory is derived from the FQ3 balance-sheet inventory of $4,523M less the $195M inventory movement reported in the FQ3 cash flow statement. FQ2 non-GAAP EPS is computed as FQ2 non-GAAP net income of $12,074M over 4,940M non-GAAP diluted shares.
Revenue
The $29.591B print beat the company's own guide by roughly $190M and the consensus range by $160M to $230M. The composition matters more than the size. AI semiconductor revenue of $16.7B was the number the FQ2 selloff was fought over, and it came in $0.7B above the $16B guide that the market treated as a disappointment three months ago. Custom XPUs were 73% of AI revenue (roughly $12.2B, up more than 3.5x year over year) and AI networking the remaining 27% (roughly $4.5B, up more than 2.5x). Both halves of the AI franchise more than doubled, which is the answer to the FQ2 worry that networking would fade as XPU volume scaled: the mix shifted, the absolute dollars did not.
The two soft spots are real and worth naming. Non-AI semiconductors landed near $4.2B against a ~$4.5B guide, growing 5% year over year and flat sequentially after management had guided 12% growth and called the cyclical recovery underway at FQ2. Infrastructure software of $8.752B (+29%) came in below the ~$8.9B guide and the +31% growth management had framed. Neither is thesis-relevant at this scale, but both are guide misses in a quarter otherwise defined by beats, and the FQ4 guides for both lines are essentially flat sequentially. The non-AI cyclical recovery that FQ2 declared has not arrived on schedule.
Margins
Non-GAAP gross margin of 75.0% beat the ~74% guide and fell 212bps sequentially, exactly as management had framed it: XPU mix with rising memory content dilutes the blended number. The correct frame is the segment level, and management supplied it, though not without stumbling. On the prepared remarks the CFO stated semiconductor segment gross margin was "approximately 76%," then corrected it mid-Q&A to approximately 67%. The corrected figure is the right one: back out the 94%-margin software segment from consolidated non-GAAP gross profit and the residual implies 67.0% for semiconductors, down from roughly 70% a quarter ago.
Operating margin is where the story lives. At 67.9% it set a record, up 237bps year over year and 63bps sequentially, against a ~67% guide. The semiconductor segment ran a 61% operating margin on operating expenses equal to 6% of segment revenue, with segment opex growing 22% against revenue up 127%. Infrastructure software ran approximately 84%, up 650bps year over year. The FQ4 guide takes operating margin to ~66% on gross margin of ~73%, so the company is explicitly signalling that it will trade five points of gross margin for a flat operating margin and 93% revenue growth. That is the right trade, and the market should stop grading Broadcom on the blended gross-margin line.
EPS and Cash
Non-GAAP EPS of $3.32 (+96%) is fully operational. The non-GAAP tax rate computes to 16.0% on the reported reconciliation, in line with the guided rate and unchanged as a driver. The diluted non-GAAP share count of 4,937M was down slightly from 4,972M a year ago. GAAP EPS of $2.68 grew 215%, more than double the non-GAAP growth rate, because the add-backs are shrinking in absolute dollars while revenue nearly doubles: acquisition-intangible amortization was $2,006M versus $2,026M a year ago, and stock-based compensation actually fell.
Free cash flow of $13.665B converted 83% of non-GAAP net income and 104% of GAAP net income, at a 46.2% margin. The working-capital detail is clean rather than flattering: days sales outstanding improved to roughly 42 days from roughly 45 in FQ2, and inventory rose only 4.5% sequentially to $4.523B after a much larger FQ2 build, even as revenue grew 33%. A company stuffing the channel does not show falling DSO and a decelerating inventory build into a 54% sequential jump in AI revenue. Capital allocation shifted materially: $3.1B of dividends, $5.6B of long-term debt repaid in the quarter with another $1.5B of senior notes retired after quarter end, and zero share repurchases.
Segment Performance
Segment Revenue Mix — FQ3 FY2026
| Segment | Revenue ($M) | % of Total | YoY Growth | Segment Gross Margin | Segment Operating Margin |
|---|---|---|---|---|---|
| Semiconductor solutions | 20,839 | 70% | +127% | ~67% | 61% (+440bps YoY) |
| Infrastructure software | 8,752 | 30% | +29% | 94% | ~84% (+650bps YoY) |
| Total net revenue | 29,591 | 100% | +86% | 75.0% | 67.9% |
Semiconductor Composition — AI vs. Non-AI
| Line | FQ3 FY26 | % of Total Revenue | YoY Growth | QoQ Growth | FQ4 FY26 Guide |
|---|---|---|---|---|---|
| AI semiconductor | ~$16.7B | 56% | +221% | +54% | $21.7B (+236% YoY) |
| of which custom XPUs (73%) | ~$12.2B | 41% | +3.5x | n/a | tripling YoY |
| of which AI networking (27%) | ~$4.5B | 15% | +2.5x | n/a | tripling YoY |
| Non-AI semiconductor | ~$4.2B | 14% | +5% | flat | ~$4.3B |
| Semiconductor solutions (reported) | $20.839B | 70% | +127% | +39% | ~$26.1B (+136% YoY) |
AI and non-AI component dollars are the rounded figures management quoted on the call and do not sum precisely to the reported segment total; the Semiconductor solutions row is the income-statement figure. XPU and networking splits are computed from the disclosed 73% XPU share of AI revenue.
AI Semiconductor: $16.7B, and the Guide the Market Punished Turned Out to Be Conservative
AI semiconductor revenue of $16.7B grew 221% year over year and 54% sequentially, and represented 56% of total revenue against 49% in FQ2. Three months ago the market sold Broadcom 13% because management guided this line to $16B rather than the roughly $17.2B the buy side wanted. The line printed at $16.7B, roughly 70% of the way to the whisper that was treated as unreachable, and the FQ4 guide of $21.7B sits above the ~$21B that FQ2's half-over-half math implied. In two quarters the company has delivered above its own framework twice.
The product detail was unusually specific. Ironwood TPU v7 shipped in high volume to both Anthropic and Google. The next-generation TPU v8i entered production shipments for Google with more memory and bandwidth than Ironwood, optimized for inference. OpenAI's first-generation accelerator, Jalapeño, shipped in the quarter. Meta's custom MTIA accelerator reaches production shipments in FQ4.
"Q3 demand was simply hot and we're just getting started."
— Hock Tan, President & CEO
Two competitive datapoints are worth extracting because they speak to the durability of the franchise rather than the size of the quarter. On the TPU roadmap, management stated that Broadcom "is now shipping the TPU v8i ahead of the MediaTek version v8t, which in fact was initiated earlier." On performance, the v8i was described as "comparable, if not surpasses the Vera Rubin GPU," and Jalapeño as outperforming "Grace Blackwell GPU for inference workloads."
Assessment: Shipping a later-initiated program ahead of a named second source is the single most load-bearing competitive disclosure of the call, because the entire multi-source bear case on Google rests on the premise that alternatives can execute on the same timeline. They have not. The performance claims against Vera Rubin and Grace Blackwell are management assertions and should be discounted accordingly, but the schedule claim is verifiable and it is the one that protects the anchor account.
AI Networking: 27% of AI Revenue and Growing as Fast as XPUs
AI networking was roughly $4.5B, up more than 2.5x year over year, and settled at 27% of AI revenue. At FQ2 management had described the ~40% networking share as a peak produced by favourable alignment, and said it would drift toward 30% as XPU volume scaled. That happened on schedule, and the important read is that the absolute dollars grew 2.5x while the share fell. Management now expects networking revenue to "grow just as fast as XPUs over the next few years," which is a stronger statement than the FQ2 framing.
The product cadence continued: Tomahawk 6 at 100 Tbps shipping in two SerDes versions, Tomahawk 7 at 200 Tbps now added, PCI Express switching leadership for scale-in, and an expanding position in optical DSPs, EMLs, VCSELs and CW lasers. Tomahawk Ultra, which enables scale-up inside a rack over low-latency Ethernet, is deploying from this quarter.
"Our Tomahawk 6 has been a phenomenal success. It started really first in scale-out, as you mentioned. It is available both in 100 Gb and 200 Gb SerDes. We actually have two versions of Tomahawk 6. Both are very successful and both are deployed in pretty much all of the AI hyperscalers that are building XPUs with us, and even those that are not using our XPUs are using the Tomahawk 6, both 100 Gb and 200 Gb."
— Charlie Kawwas, President, Semiconductor Solutions Group
Assessment: The disclosure that Tomahawk 6 is deployed at hyperscalers who are not Broadcom XPU customers is the part that matters. It decouples the networking franchise from the custom-silicon franchise and means AI networking is not an attach-rate business dependent on winning the accelerator socket. That is a materially better structure than the market credits, and it is why we keep Bull 2 intact even as the networking share of AI revenue declines.
Non-AI Semiconductors: The Cyclical Recovery Stalled
Non-AI semiconductor revenue of approximately $4.2B was "up 5% year-on-year and flat sequentially," with broadband and server storage up and wireless down. FQ2 guided this line to ~$4.5B and +12%, and framed bookings above $6B as clear evidence that a full cyclical recovery was under way. The FQ4 guide of ~$4.3B implies the recovery continues to be gradual rather than the acceleration described three months ago.
Assessment: This is the clearest failed commitment from the FQ2 call. At 14% of revenue it does not move the thesis, and the wireless decline is seasonal and known. But it does remove the "second engine" argument that FQ2 leaned on, and it means the growth story is now almost purely AI. Investors who bought the diversification framing at FQ2 should mark it down.
Infrastructure Software: 29% Growth, ARR Decelerating
Infrastructure software revenue of $8.752B grew 29% year over year against a ~$8.9B and +31% guide, with gross margin of 94% and operating margin near 84%, up 650bps year over year. Annual recurring revenue growth was 15%, down from the 17% cited at FQ2. The FQ4 guide of ~$8.7B is flat to slightly down sequentially and represents +25% year over year, a further deceleration.
The product news was the VMware Private AI Cloud announcement, positioned as a secure enterprise platform for building and running AI alongside existing applications, plus continued VMware Cloud Foundation traction in repatriating workloads from public cloud to private.
"Enterprise consumption of AI is in fact opening a new opportunity for our infrastructure software business."
— Hock Tan, President & CEO
Assessment: Software is doing exactly what it is supposed to do, which is throw off 84%-margin operating income and stabilize consolidated cash flow while the semiconductor business does the growing. The ARR deceleration from 17% to 15% and the guided step-down to +25% mark the end of the VMware repricing tailwind; from here this is a mid-teens ARR business with elite margins, not a growth story. That is fine, but nobody should model software as a source of upside surprise.
Key Topics & Management Commentary
Overall Management Tone: Materially more forward-leaning than FQ2, and deliberately so. Where the prior call reiterated frameworks and declined to be drawn into higher headline numbers, this one raised the current year, raised the next year by roughly 15%, introduced a year that did not previously exist, and attached an earnings figure to it. The posture on demand was unhedged; the posture on the financing structures that underwrite part of that demand was the opposite, with the only flatly non-responsive answer of the call coming when management was asked for an aggregate exposure number. Analyst pushback was concentrated, competent and almost entirely about arithmetic rather than strategy, and management engaged the arithmetic openly enough to concede that not all of the gigawatts it counted will actually deploy.
1. The 0.7% That Moved the Stock
The FQ4 revenue guide of approximately $34.8B landed below a consensus range of $35.03B to $35.1B, with some estimates above $36B. That is a shortfall of roughly $230M to $300M on a $34.8B number, or 0.7% to 0.9%. It arrived alongside a raised full-year AI number, a raised FY27 number, a brand-new FY28 number and a first-ever out-year EPS marker. The stock fell mid-single digits after hours.
The mechanics are the same as FQ2 and worth stating plainly: when a stock is priced to require an upward revision on the near-quarter line every ninety days, the near-quarter line becomes the only line that trades. The composition of the guide is arguably stronger than the headline, since AI is guided to $21.7B (above the ~$21B that FQ2's own math implied) and the shortfall sits in the two slow lines, non-AI semis and software, that together are 38% of FQ4 revenue and 0% of the investment case.
Assessment: A 0.7% guide shortfall driven by the non-AI half of the business, in the same release that raised the AI framework by 15% for FY27 and put $230B on FY28, is not new information about Broadcom. It is information about positioning. This is the second consecutive quarter the tape has traded the near-term line against a strengthening multi-year one, and the second consecutive quarter the resulting price is better than the one before it.
2. FY26 Raised to $58B, FY27 Raised to $115B, FY28 Introduced at $230B
Management raised the FY26 AI revenue framework to $58B, up 186% year over year and above the prior ~$56B guidance. FY27, which stood at "in excess of $100B" at FQ2 and was pointedly not raised then, moved to approximately $115B. FY28 was introduced at approximately $230B. Both out-years were described as supply-secured, which is a stronger claim than demand visibility.
"In 2027, we have secured the supply to again double AI revenue to approximately $115 billion. Our demand actually exceeds this outlook, and we will work to improve supply."
— Hock Tan, President & CEO
"We have line of sight for fiscal 2028 AI semiconductor revenue growth to again double to $230 billion. Here again, we have secured the supply to meet this outlook."
— Hock Tan, President & CEO
The framing was also explicitly de-linked from the quarterly cycle, which matters for how the next three prints get read.
"This AI revenue guidance is being provided to you to give you the trajectory of our growth, and we do not intend to update it on a quarterly basis."
— Amie Thuener, CFO
Assessment: The FQ2 recap named a quantified backlog or an FY27 raise as the explicit watch item for this print. The FY27 raise arrived and was larger than expected. The distinction between "secured the supply" and the more usual "visibility" language is the substantive part: management is claiming the constraint has moved from whether it can sell the chips to whether the customers can energise the buildings. That is a materially different risk profile from a demand forecast, and it is why we treat FY27 as high-confidence and FY28 as a framework.
3. The First Out-Year EPS Marker: More Than $30 in FY28
Buried at the end of the AI outlook paragraph, management attached an earnings number to the revenue framework for the first time.
"As a result, I have to say, we are very much on target to exceed $30 in earnings per share in fiscal 2028."
— Hock Tan, President & CEO
The number is internally coherent with the rest of the guidance. Take FY28 AI revenue of $230B, add non-AI semiconductors and infrastructure software at roughly their current run-rates, apply the 66% operating margin the company is already guiding to for FQ4, subtract interest at the current run-rate and tax at the guided 16%, and divide by a roughly 5.0B non-GAAP share count. The arithmetic lands in the low thirties. This is not a stretch target bolted onto the revenue framework; it is the revenue framework carried through an unchanged margin and tax structure.
Assessment: This changes what the stock is being valued against. At the $367.24 pre-print close, AVGO trades near 32x our FY26 non-GAAP EPS estimate of roughly $11.61 and roughly 12x management's own FY28 marker. Put differently, at a market multiple of about 22x the current price discounts roughly $16.70 of FY28 EPS, which is a 20% two-year earnings CAGR against a company framework calling for roughly 61%. The market is paying for a little over half of what management just said. That gap is the entire investment case, and it is also the reason the rating does not need the $230B to be exactly right.
4. Anthropic Displaces Google as the Largest XPU Customer
The most consequential structural disclosure of the call was the inversion of the customer hierarchy. Anthropic starts from 1 GW of Ironwood deploying in 2026, adds 5 GW of TPU v8i in 2027, and carries a further 10 GW of line-of-sight in 2028.
"Anthropic is on track to become our largest XPU customer in 2027 and sustain that in 2028."
— Hock Tan, President & CEO
"In 2028, we have clear line of sight to deliver an incremental 10 GW, even as we expect Google to grow for us."
— Hock Tan, President & CEO
OpenAI follows at 1.3 GW of Jalapeño in 2027 and more than 5 GW in 2028 across Jalapeño and its successor, making it the second-largest XPU customer. A third-generation OpenAI XPU is already in development, and the second generation is approaching tape-out. Meta takes three generations of MTIA between now and the end of 2027, totalling 3 GW through 2028.
Assessment: A decade of Broadcom's custom-silicon franchise has been the Google TPU program. In two years the largest customer becomes a private company that does not yet generate the cash flow to fund its own deployments, and the second-largest becomes another one. That is simultaneously the bull case (the franchise has broadened beyond a single hyperscaler and is winning the frontier labs on merit) and the sharpest possible statement of the bear case (the incremental revenue is concentrated in counterparties whose capacity to pay depends on capital markets rather than operating cash flow). Bull 1 is confirmed; Bear 2 has changed shape rather than resolved.
5. Google Reframed as Multi-Tens of Billions of Dollars of TPUs Annually
At FQ2 management would only characterise the Google long-term agreement as "very, very substantial" in dollars while conceding Google would multi-source. This quarter it put a magnitude on it.
"Under this agreement, we're planning to deliver multi-tens of billions of dollars of TPUs annually over the next several years."
— Hock Tan, President & CEO
Management attributed the durability of the relationship to a specific technical stack rather than a commercial one: SerDes, chip-to-chip interconnect, HBM and SRAM integration, advanced packaging, and consistently fastest time to market "without the need for respins." Demand growth from Google was pushed out to 2028 and 2029, fulfilled through successive TPU generations already in development.
Assessment: A figure of "multi-tens of billions of dollars of TPUs annually" against FY26 AI revenue of $58B implies Google alone is a meaningfully large share of the current business and stays large in absolute terms while ceding the top spot. That is the right way to read it: Google is not shrinking, it is being out-grown. The specificity is also a step forward in disclosure, and it partially answers the FQ2 watch item on Google share without ever giving a share number.
6. Broadcom Is Now Buying Its Own Bottlenecks
Capital expenditure was $532M in the quarter, up 275% year over year, and is guided to $1.4B in FQ4. That takes FY26 capex to roughly $2.4B against $386M in the first three quarters of FY25. The money is going into two specific bottlenecks: substrates and lasers.
"Well, we are going to start deploying our Singapore fab for substrates, by the way, starting fiscal 2027. That would, I guess, address a key part of our supply bottlenecks."
— Hock Tan, President & CEO
"But also our EML, CW and VCSEL factories, our indium phosphide factories, both in the U.S. as well as in Singapore, we're actually more than tripling them year-on-year."
— Charlie Kawwas, President, Semiconductor Solutions Group
Management was explicit that laser demand "is far surpassing supply out there in the industry," and framed the expansion as enabling the ecosystem rather than purely capturing margin.
Assessment: This is a genuine change in the business model and it is being underplayed. Broadcom has run a fab-lite structure for its entire post-Avago history; committing owned capacity in substrates and indium phosphide is a deliberate decision to internalise the two constraints that could cap the FY27 and FY28 frameworks. At $2.4B it is still trivial against $54B of annualised free cash flow, so the capital cost is not the point. The point is that the supply claim behind the $115B and $230B numbers has an owned-asset component, which makes it more credible than a purchase commitment would be.
7. The XPV Platform: What Was Disclosed
The financing vehicle announced at FQ2 now has a name and a structure. The AI XPV Platform, built with Apollo and Blackstone, is designed to enable more than 20 GW of compute infrastructure for OpenAI and Anthropic by the end of 2028. The first $35B tranche closed in June for Anthropic's 1 GW deployment, which is already underway. Management laid out a three-part rationale: bridging the gap between the labs' current cash flow and their upfront investment needs, enabling execution where demand is locked in, and reviewing each financing through both a commercial and a balance-sheet lens.
"Through the XPV Platform, we partner with sophisticated third-party financial partners to independently underwrite and capitalize the assets, rather than providing the direct financing ourselves. Where necessary, we may provide modest residual value guarantees, which are contingent liabilities we view as low risk, supported by the strong profitability trajectory of these labs and the sustaining value of the underlying assets."
— Amie Thuener, CFO
The economic logic management offered for why this is a good trade rather than a favour was blunt.
"It is a great investment for us when you think about it, that every gigawatt of compute they deploy, they could achieve $30 billion of ARR, annual revenue per gigawatt. That is a hell of a business model."
— Hock Tan, President & CEO
Assessment: The structure is better than the FQ2 description implied. Third parties underwrite and capitalise; Broadcom's participation is a residual value guarantee rather than principal. The characterisation of the guarantees as "modest" and "low risk" is management's, and it rests on an assumption about the profitability trajectory of two companies that do not publish financial statements. What was disclosed is a framework. What was not disclosed is the number, which is the next topic.
8. The XPV Platform: What Was Not Disclosed
Asked directly for the maximum off-balance-sheet risk, an analyst put the prior 10-Q's figure on the record and management confirmed it while refusing to extend it.
Q: "Can you shed a little bit more light on the maximum off-balance sheet risk for the backstop agreements? Your last Q showed that the first tranche had maximum exposure of about $29 billion. Is this about the order of magnitude for each of the Anthropic and OpenAI gigawatts over the next few years?"
— Jack Adair (for Ben Reitzes), Melius Research
A: "Thanks, Jack. Listen, we don't have anything to announce today on residual value guarantees or backstop. There's nothing new to add. The numbers that you outlined around what we've already done remain true."
— Amie Thuener, CFO
The arithmetic that management declined to do is straightforward and unflattering. Roughly $29B of maximum exposure attaches to one gigawatt. That is 29% of Broadcom's $99.69B of book equity, against a single deployment, and the platform is designed for more than 20 GW. Management was clear that not every tranche will carry Broadcom participation, that four of the six XPU customers self-fund, and that future structures will differ. Linear extrapolation is therefore wrong. But the direction of the bound is not in question, and investors currently have no disclosed ceiling of any kind.
Assessment: This is the reason the thesis carries a live risk rather than a watch item. The disclosure asymmetry is stark: management supplied gigawatt-by-customer detail through 2028, a two-year revenue total, and an EPS figure, then declined to supply a single aggregate number on contingent exposure. A company willing to forecast $230B of revenue two years out is capable of bounding its own guarantee book. Until it does, position sizing has to account for a contingent liability that could plausibly reach a material fraction of equity, and the quarterly 10-Q footnote is now the most important page in the filing.
9. Management Conceded the Gigawatts Will Not All Deploy
Pressed on whether the disclosed gigawatt counts reconcile to the revenue framework, management did something unusual for a company mid-upgrade: it argued down its own number.
"What we're saying is, as you add up the gigawatts, we're not saying that over the next two years, 2027, 2028, that there are 30 GW that will go into production, and therefore we ship it. We think we judge it conservatively to be somewhat less."
— Hock Tan, President & CEO
The reconciliation offered was a two-year aggregate rather than a gigawatt count: "we believe with a pretty high degree of confidence, we will ship $350 billion of AI semiconductors to these customers in the next two years." That total is the sum of the $115B and $230B outlooks and implies roughly $11.5B of Broadcom content per gigawatt across 30 GW.
Assessment: The concession is credibility-positive and estimate-negative at the same time, which is the correct combination. Management is telling investors that the gigawatt disclosures are a demand map, not a shipment schedule, and that the revenue outlook already contains a haircut. That is a more honest construction than most AI-exposed guidance, and it is also a warning that anyone building up from the gigawatt table will overshoot.
10. Content Per Gigawatt Is Stable, Not Rising
At FQ2 management argued content per gigawatt would rise generation over generation as XPUs added SRAM, embedded CPU cores and more HBM. This quarter the framing changed.
"So what we are seeing per gigawatt is in the range of less than $30 billion, $20 billion-$30 billion per gigawatt. And we expect that to be very sustaining in that level because the power of each chip goes up. So even as you increase the price of the chip, the content in dollars is relatively stable."
— Hock Tan, President & CEO
The mechanism is that higher-performance accelerators draw more power, so fewer of them fit inside a gigawatt; the higher unit ASP is offset by lower unit volume per gigawatt. Note that the $20B to $30B figure describes total infrastructure content, not Broadcom's own take, which the $350B-over-30GW arithmetic puts nearer $11.5B.
Assessment: This is a quiet downgrade of one of the FQ2 bull arguments and should be recorded as such. Growth is now a pure gigawatt-count story, with content per gigawatt flat rather than compounding. It does not change the FY27 or FY28 numbers, which are given directly, but it removes an ASP tailwind from the FY29-and-beyond extrapolation that some models carry.
11. Zero Buybacks in a Quarter the Stock Fell 23% From Its Last Print
Broadcom repurchased no shares under its repurchase program in FQ3, having bought $7.85B in FQ1 and $0.6B in FQ2. It instead retired $5.6B of long-term debt during the quarter and a further $1.5B of senior notes after quarter end, leaving gross principal fixed-rate debt at $59.6B with a 4% weighted-average coupon and 7.4 years to maturity. Dividends of $3.1B were paid. The stock closed at $367.24 entering this print, 23% below the $479.23 close entering the last one.
The topic was not raised in the prepared remarks beyond the factual capital-allocation recap, and no analyst asked about it.
Assessment: There are two readings and they point in opposite directions. The benign one is that a company about to step up capex, sitting on a growing contingent-guarantee book and carrying $35.4B of net debt should deleverage before it buys stock, and $7.1B of debt retirement across two months is exactly that. The uncomfortable one is that a management team publicly forecasting more than $30 of FY28 EPS declined to buy its own shares at 12x that number. We lean to the first reading given the XPV exposure, but the silence on both sides of the call is itself notable.
12. Six Customers, and Management Is Glad There Are Only Six
Asked to enumerate the constraints on the FY27 outlook, management gave the fullest answer of the call, listing leading-edge silicon availability, substrates, HBM, system memory that Broadcom does not even supply, and land, power and shell.
"All these are multidimensional problem coming from various sources, depending on different times, each might become a bottleneck. It's a constant interesting challenge as we work this through, which is why, in some ways, I'm so glad we have only six customers to deal with."
— Hock Tan, President & CEO
On the deployment gate specifically, management described analysing customer readiness before crediting demand: "We are very engaged with each of them on how much LPS, land, power, and systems they have before we actually believe it will happen."
Assessment: The remark about being glad to have only six customers is the most honest sentence on the call. Concentration is being described as an operational advantage, which in a supply-constrained ramp it genuinely is, and it is simultaneously the structural vulnerability of the equity. Both things are true. The LPS diligence process is the right answer to the question of whether the outlook is a customer wish-list, and it is consistent with management having already haircut the 30 GW.
Guidance & Outlook
FQ4 FY2026 Guidance
| Metric | FQ3 FY26 Actual | FQ4 FY26 Guide | YoY Growth Implied | Change vs. Street |
|---|---|---|---|---|
| Total revenue | $29.591B | ~$34.8B | +93% | -0.7% to -0.9% |
| Semiconductor solutions | $20.839B | ~$26.1B | +136% | Above |
| of which AI semiconductor | ~$16.7B | $21.7B | +236% | Above FQ2-implied ~$21B |
| of which non-AI semiconductor | ~$4.2B | ~$4.3B | low single digit | Below |
| Infrastructure software | $8.752B | ~$8.7B | +25% | Below |
| Non-GAAP gross margin | 75.0% | ~73% | -500bps YoY (from 78%) | Lower on XPU memory mix |
| Non-GAAP operating margin | 67.9% | ~66% | flat YoY | Maintained |
| Non-GAAP tax rate | 16.0% | ~16% | n/a | Maintained |
| Non-GAAP diluted share count | 4,937M | ~4,940M | n/a | Ex-buybacks |
| Capital expenditures | $532M | ~$1.4B | n/a | Capacity build |
Multi-Year AI Framework
| Period | Prior Framework (FQ2, June 2026) | New Framework (FQ3) | Change | Basis |
|---|---|---|---|---|
| FY2026 AI semiconductor revenue | ~$56B | $58B | Raised ~4% | +186% YoY; FQ4 guide of $21.7B |
| FY2027 AI semiconductor revenue | "in excess of $100B" | ~$115B | Raised ~15% | Supply secured; demand exceeds outlook |
| FY2028 AI semiconductor revenue | not given | ~$230B | New | Supply secured |
| FY2027 + FY2028 combined | not given | ~$350B | New | Across ~30 GW of demand, judged conservatively lower on deployment |
| FY2028 non-GAAP EPS | not given | >$30 | New | First explicit out-year EPS marker |
The near-quarter guide and the multi-year guide point in opposite directions, which is the entire reason the tape and the fundamentals disagreed. FQ4 revenue of ~$34.8B requires 17.6% sequential growth on top of a quarter that already grew 33.4% sequentially, and takes AI to 62% of total revenue. The gross-margin guide of ~73% is 500bps below the year-ago quarter, and management pre-empted the question by telling investors to stop grading the line.
"It is not the first time we have told you guys that, because stop focusing on gross margin is what we are saying. Look at where it matters, operating margin at the end of the day, because the growth in revenue far surpasses their growth in OpEx, operating spending, to support their growth in revenue."
— Hock Tan, President & CEO
Implied quarter-over-quarter ramp: The FY26 AI framework of $58B against the first three quarters (roughly $8.4B, $10.8B and $16.7B) requires an FQ4 of roughly $21.7B, which is exactly what was guided. The framework and the quarterly guide are the same statement. For FY27, reaching ~$115B from an FQ4 exit rate of $21.7B requires an average quarterly AI run-rate near $28.75B, or roughly 33% above the FQ4 exit. Management characterised 2027 deployments as back-half loaded at FQ2 and did not revise that shape, so the implied FY27 exit rate is well above $30B per quarter.
Street at: Consensus for FQ4 sat at $35.03B to $35.1B and will now be marked to the $34.8B guide, with the reduction concentrated in non-AI semiconductors and software. FY27 and FY28 estimates move the other way and by a much larger absolute amount: the AI line alone rises roughly $15B for FY27 against the prior "in excess of $100B" anchor, and FY28 gains a company-sanctioned number where most models carried an extrapolation. The net revision to two-year earnings power is sharply positive even though the next-quarter revision is negative.
Guidance style: Management guides to contracted, supply-secured volume and says so explicitly. Asked what could push the FY27 number higher, the answer was a description of deliberate conservatism rather than a stretch case: "We are very careful, and to be honest, we try to be conservative." The company also volunteered that it has already haircut the gigawatt count for deployment risk. Across the last two prints this team has guided AI below the buy-side whisper and then shipped above its own guide, which is the pattern to extrapolate.
Analyst Q&A Highlights
What Actually Caps the FY27 Number
The opening question went straight to the structural issue behind a doubled outlook: if demand exceeds the guide, what is the binding constraint, and what would have to change to lift it. Management's answer separated two things that are usually conflated, namely the ability to ship chips and the customer's ability to deploy them, and located the conservatism in the second.
Q: "You talked about the business doubling next year, and you said demand could be higher than that. Can you talk about the supply around that and what are the supply bottlenecks, and what are the variables that could drive that number higher if you are able to resolve them?"
— Joseph Moore, Morgan Stanley
A: "Yeah, demand, we can ship significantly more. Question in our mind sometimes is, even as we ship the chips, are they going to be deployed on a timely basis? That's always very much in our mind when we give you that outlook. But certainly our customers want us to ship more. But we have secured supply, and we think it's the right number to put it to 115."
— Hock Tan, President & CEO
Assessment: The most useful answer of the call, because it reframes the risk. If supply is secured and demand exceeds the outlook, then the $115B is a deployment forecast, not a sales forecast, and the variance driver is data-centre readiness at the customer rather than anything inside Broadcom. That is a risk investors can monitor externally through power interconnect queues and construction announcements, which is more than can usually be said for a semiconductor outlook.
Reconciling the Gigawatt Table to the Revenue Framework
The sharpest question of the call did the arithmetic in public: sum the disclosed gigawatts, divide the revenue outlook by them, and note that the resulting content per gigawatt is a fraction of what a competitor claims. Management conceded both the gigawatt total and, more importantly, that the revenue outlook already assumes not all of it deploys.
Q: "If it is, if I calculate the revenue per gigawatt, I come out with your revenue guides, I come out somewhere between $11 billion and $12 billion per gigawatt. Your competitor suggested something like 40. I'm just wondering, is this the right level that we should think about for your content, and how should that trend as we go forward, as you said, into more generations of XPUs, again, with higher performance and higher memory?"
— Stacy Rasgon, Bernstein
A: "We are giving you a judged number of $115 billion of chips to these guys in fiscal 2027, another $230 billion in 2028, which would add up, I know, to about $350 billion. Another way of saying is, we believe with a pretty high degree of confidence, we will ship $350 billion of AI semiconductors to these customers in the next two years. That's the best way to look at it. It doesn't necessarily mean all 30 GW need to have been deployed within that period."
— Hock Tan, President & CEO
Assessment: Management substituted a dollar commitment for a gigawatt commitment, which is the right unit for an investor and a deliberate narrowing of what it is willing to be held to. The content-per-gigawatt gap versus the merchant-GPU comparison is not a weakness; it is the product of selling silicon rather than full systems, and it is the same gap that lets the customer claim half the cost. Anyone modelling from the gigawatt table upward will overshoot the revenue line, and management said so.
The Aggregate Ceiling on Off-Balance-Sheet Exposure
The only question management flatly declined to answer concerned the size of the guarantee book. The exchange is short and the refusal is unambiguous, which is why it is reproduced in full rather than characterised.
Q: "Can you shed a little bit more light on the maximum off-balance sheet risk for the backstop agreements? Your last Q showed that the first tranche had maximum exposure of about $29 billion. Is this about the order of magnitude for each of the Anthropic and OpenAI gigawatts over the next few years? Can you add a little bit more color as well on the residual value and the risk associated?"
— Jack Adair (for Ben Reitzes), Melius Research
A: "I can't give you an overarching look at what's the max and what each one's going to look like. But we will tell you at the right time. We have nothing to announce today."
— Amie Thuener, CFO
Assessment: A company that will forecast revenue two years out and attach an EPS figure to it is choosing not to bound its own contingent liabilities. The refusal was procedural rather than evasive in tone, and the deal-by-deal reasoning offered alongside it is legitimate. But the asymmetry between what is forecast and what is withheld is now the single largest gap in the disclosure, and it will define how the equity trades every time credit conditions for private AI labs tighten.
Whether the Financed Labs Actually Control Their Own Silicon Choice
A pointed challenge argued that frontier labs depend on land, power and shell supplied by cloud partners and funding partners with their own silicon preferences, so a multi-year commitment from a lab may not be the lab's to make. Management's answer was that the labs are becoming hyperscalers themselves, and that the current dependence on third-party capacity is transitional.
Q: "In many cases, these frontier labs depend on land, power, shelf from their cloud service providers, sometimes NVIDIA-related entities. How much of their use of silicon is truly their choice versus what their CSP or funding partner dictates?"
— Vivek Arya, Bank of America
A: "Long term, we see these guys to be no different from a hyperscaler and will run their own data centers and be first party to offer AI, generative AI, APIs, as SaaS models to the world. We see that happening. It is not speculation, it is actually happening, and we are in the midst of enabling that."
— Hock Tan, President & CEO
Assessment: The answer is a forecast, not evidence, and it should be held as such. The load-bearing claim is that first-party compute is the destination for both labs, which if true makes the silicon decision genuinely theirs and makes the multi-year commitments durable. If it is not true, or is slower than assumed, the 2028 gigawatts are exposed to counterparties who do not control their own procurement. This is the least-supported link in an otherwise heavily-evidenced outlook.
Where Consolidated Gross Margin Settles as Memory Content Rises
The same questioner asked for a multi-year gross-margin path given the XPU mix and rising memory cost. The answer contained the call's only self-correction, restating a prepared-remarks figure that had been wrong by nine points, and then declined the multi-year framing entirely in favour of a one-quarter horizon.
Q: "Amie, I just wanted to clarify any impact on gross margins in 2027 and 2028, given this XPU mix and rise in memory cost."
— Vivek Arya, Bank of America
A: "First, I want to just make sure I correct what I said before, just to be super crisp. Gross margin for our Semiconductor Solutions segment was approximately 67% in Q3. To your question, gross margin really reflects revenue mix between both semiconductors and infrastructure software. Then within semiconductors, the product mix is reflected, as well as increasing memory content. As you can see, as XPUs become a larger proportion of our total revenue mix, it impacts our margin. We guide one quarter at a time, so we will tell you each quarter what our margin is going to look like."
— Amie Thuener, CFO
Assessment: The corrected 67% is verifiable from the release: strip the 94%-margin software segment out of consolidated non-GAAP gross profit and the semiconductor residual computes to 67.0%. The correction was the right call and made promptly, though a nine-point error in prepared remarks on the single most-scrutinised margin line is a process observation worth carrying forward for a CFO in her second quarter. The substantive answer, a refusal to give any multi-year gross-margin path while giving a two-year revenue path, is consistent with the operating-margin-only framing management is pushing.
Ethernet in Scale-Up and Whether Networking Rides the Accelerator Socket
A recurring line of questioning probed whether the networking franchise is an attach-rate business dependent on winning the custom-silicon socket, and how far the low-latency Ethernet scale-up product has penetrated. The answer disclosed deployment at accounts that are not Broadcom XPU customers, which is the decoupling investors have wanted.
Q: "Out of your six frontier model-based XPU customers, how many of them are using Tomahawk scale-out networking? Can you also just give us an update on the adoption of your Tomahawk Ultra scale-up platform as well?"
— Harlan Sur, JPMorgan
A: "We actually have two versions of Tomahawk 6. Both are very successful and both are deployed in pretty much all of the AI hyperscalers that are building XPUs with us, and even those that are not using our XPUs are using the Tomahawk 6, both 100 Gb and 200 Gb. With respect to Tomahawk Ultra, we have quite innovated ahead of the market here by enabling scale-up using low latency Ethernet. The adoption also on this device has surprised us, and we are starting to see it deployed, starting actually this quarter and in FY 2027 coming up in scale-up applications."
— Charlie Kawwas, President, Semiconductor Solutions Group
Assessment: The phrase that matters is "even those that are not using our XPUs are using the Tomahawk 6," and it was volunteered rather than extracted. It means the networking franchise has a total addressable market that includes merchant-GPU clusters, which is a structurally larger and less concentrated business than the custom-silicon franchise it sits alongside. The Tomahawk Ultra adoption surprising management to the upside adds a second scale-up vector that was not in the FY27 framework a quarter ago.
What Actually Gates Deployment Beyond Land, Power and Shell
A closing line of questioning asked management to rank the constraints and to say which could be relieved. The answer named five distinct bottlenecks across three tiers of the supply chain, including one Broadcom does not even supply, and framed them as rotating rather than static.
Q: "I was wondering if you could maybe address some of the other constraints that are impacting your outlook. You mentioned land, power, shell. Do you think that is the largest of the constraints that you face heading into fiscal 2027? Specifically, can you address any other supply chain constraints, whether it be memory, substrates, or other components that could be impacting your outlook, and to what extent you could expect those to get ameliorated?"
— Jim Schneider, Goldman Sachs
A: "Then even deeper than that, we talk about substrates, which is a specific issue, which is leading us to build our own substrate capacity at scale in our factory in Singapore, jointly with one of our partners. Of course, we all know about memory, HBM memory. Beyond HBM memory, the system memory that goes into AI servers, which we don't supply necessarily, but our customers have to secure too."
— Hock Tan, President & CEO
Assessment: The candour about system memory, a component Broadcom neither sells nor controls, is the most useful risk disclosure in the transcript. It means the FY27 and FY28 outlooks are exposed to a DRAM market that is currently tight and that no amount of Broadcom capacity investment can relieve. The substrate answer cuts the other way and is why the Singapore capex matters: management identified a bottleneck and is buying its way out of it rather than forecasting through it.
What They're NOT Saying
- Any aggregate cap on XPV guarantee exposure: The first tranche carries roughly $29B of maximum exposure for one gigawatt, management confirmed those figures still hold, and then declined to bound a platform designed for more than 20 GW. A company forecasting $230B of revenue two years out is choosing not to bound a contingent liability that is already 29% of book equity for a single deployment.
- A quantified backlog, for the second consecutive quarter: FQ2 disclosed more than $30B of AI bookings against $10.8B shipped and refused to confirm whether the 18-month backlog approached $200B. This quarter the bookings figure was not given at all. The multi-year revenue framework replaced it, which is arguably more useful, but the one hard number that would let investors check the framework independently has now been withheld twice.
- Revenue concentration across the six customers: Gigawatts are disclosed per customer through 2028. Dollars are not. With the largest customer changing identity in 2027 and the top two becoming private companies, the revenue split is the disclosure that would matter most and it is the one that is systematically absent.
- Why no shares were repurchased: Zero buybacks in a quarter that opened 23% below the prior print's starting price, from a management team that simultaneously forecast more than $30 of FY28 EPS. Neither the prepared remarks nor any analyst addressed the change from $7.85B of repurchases in FQ1.
- Any multi-year gross-margin path: A two-year revenue outlook and a two-year EPS marker were given. Gross margin is guided one quarter at a time, with an explicit instruction to stop looking at it. The FQ4 guide of ~73% is 500bps below the year-ago quarter, and the direction beyond that is acknowledged as down without a floor.
- The economics of Broadcom's participation in XPV: Whether Broadcom receives fees, equity, warrants or preferential supply terms in exchange for the residual value guarantees was not disclosed. The guarantees were characterised as "modest" and "low risk"; the compensation for providing them was not characterised at all.
- Apple: A July 8-K disclosed expanded multi-year agreements to supply custom ASIC silicon across multiple generations of Apple products through 2031. Management referred to six XPU customers and walked through four of them by name. The seventh anchor relationship signed during the quarter went unmentioned on the call.
- What happens to the FY28 framework if a lab misses a financing round: The two largest 2028 customers are private, unprofitable and dependent on external capital. Management addressed the mechanism that funds them and the destination it believes they are heading toward, but not the downside case. The word used for the deployment risk was "timely," never "cancelled."
Market Reaction
- Pre-print setup: AVGO closed at $367.24 on September 2, up 6.1% year to date against the S&P 500's 12.0%, up 23.1% over trailing twelve months, and down 6.4% over the trailing thirty days. The 52-week closing range entering the print was $293.41 to $481.57, putting the stock 24% below its 52-week closing high and 23% below the $479.23 close that preceded the FQ2 print in June.
- After-hours move: Shares fell as much as roughly 6% in extended trading, trading below $350 at the lows, and pared the decline to the low single digits later in the session. The move was attributed to the FQ4 revenue guide of ~$34.8B landing below the $35.03B to $35.1B consensus range, not to the reported quarter, which beat on revenue, adjusted EPS and AI revenue.
- Next-day session: The September 3 regular session had not opened at the time of publication. Levels above are the FMP-verified pre-print close and the extended-hours range.
This is the second consecutive quarter in which Broadcom has been sold on the near-term line while the multi-year line improved, and the mechanism is identical. In June the stock fell 13% because a $16B AI guide was not a $17.2B AI guide. In September it fell mid-single digits because a $34.8B revenue guide was not a $35.0B revenue guide. In both cases the company beat the quarter, and in this case it also raised the current year, raised the following year by 15% and introduced a year that carried a company-sanctioned EPS figure.
What is different this time is that the stock enters the reaction 23% lower than it entered the last one. A de-rating that began as a positioning washout in June has persisted through a quarter in which the contracted framework expanded twice, leaving AVGO up 6.1% year to date against a market up 12.0%. The AI bellwether has underperformed the index in a year its AI revenue tripled. Whatever the market is discounting, it is not the revenue trajectory.
The more defensible reading of the after-hours move is that it is about the financing question rather than the guide. The 0.7% shortfall is a rounding error against a framework that moved by tens of billions; the refusal to bound XPV exposure is not. A market that has spent 2026 learning to distrust circular financing structures in AI has a legitimate reason to discount a supplier whose two largest future customers are being funded, in part, by guarantees the supplier will not size. The guide is the stated reason. The disclosure gap is the better one.
Street Perspective
Debate: Is the FY28 $230B Framework a Contract or an Aspiration?
Bull view: The bull case being made on the Street is that management described both out-years as supply-secured rather than demand-visible, disclosed gigawatt commitments customer by customer, and volunteered a haircut on deployment before anyone asked for one. Sell-side work published earlier in 2026 had already converged toward roughly $200B of 2028 AI revenue independently, so the company number is an extension of an existing trajectory rather than a new claim.
Bear view: The bear camp contends that $230B in fiscal 2028 requires roughly twenty gigawatts to energise on schedule, that management itself conceded not all of the counted capacity will deploy, and that the two customers carrying the largest incremental commitments are private companies with no published financials and no operating cash flow to fund them. A framework whose largest single input is the capital-raising success of two counterparties is not a contract regardless of how the supply side is described.
Our take: Both are right about different years. FY27 at $115B is supply-secured against customers with contracts in hand and an FQ4 exit rate of $21.7B, and we underwrite it. FY28 at $230B is a demand map with a stated haircut, and it should be treated as a directional framework rather than an estimate. The useful consequence is that the investment case does not need FY28 to be right: at the pre-print price the market is discounting roughly half of the FY28 EPS marker, so a substantial miss against the framework still leaves the stock reasonably valued.
Debate: Does the XPV Platform De-Risk Demand or Manufacture It?
Bull view: The structure is genuinely third-party. Apollo and Blackstone underwrite and capitalise the assets; Broadcom's participation is limited to residual value guarantees on hardware with a long useful life, and four of the six XPU customers self-fund entirely. Removing the capital bottleneck for two supply-constrained labs converts locked-in demand into shipped revenue, and the economics are compelling if a gigawatt of compute genuinely supports the annual recurring revenue management described.
Bear view: A growing consensus among skeptics holds that vendor-adjacent financing flatters growth on the way up and amplifies the unwind on the way down. Roughly $29B of maximum exposure attaches to one gigawatt, the platform targets more than twenty, and management refused to give any aggregate figure. Guarantees characterised as low risk on the strength of unpublished lab profitability are precisely the exposure that gets repriced fastest when AI credit conditions turn.
Our take: We moved from cautiously bullish on this structure at the last print to genuinely concerned at this one, and the reason is the disclosure rather than the structure. The design is defensible and the third-party underwriting is real. But a company willing to guide revenue two years out and attach an EPS number to it can bound its own guarantee book, and choosing not to is a decision, not an oversight. This is now the live risk in the thesis and the item that would cause us to cut the rating if the next 10-Q shows exposure scaling without a stated ceiling.
Debate: Is the Customer Hierarchy Inversion a Broadening or a Concentration?
Bull view: Some desks argue the franchise has decisively outgrown its dependence on a single hyperscaler. Google remains enormous at multi-tens of billions annually and is still growing, while Anthropic, OpenAI and Meta add three independent multi-gigawatt programs across multiple accelerator generations. The networking business compounds the diversification, since Tomahawk 6 is deployed at hyperscalers that buy no Broadcom accelerators at all.
Bear view: The bear reading is that "six customers" is not diversification when four of them are described as the ones that are "just simply going to be huge," and when the incremental growth from 2027 onward is concentrated in two private labs. The anchor account of the last decade is being demoted by a counterparty that did not exist as a customer three years ago. Management itself described the concentration as an operational convenience.
Our take: The bull case is stronger on the technology and the bear case is stronger on the credit. Broadcom is clearly winning these sockets on execution, and shipping the TPU v8i ahead of a later-initiated second source is the strongest evidence yet that the moat is real. But the composition of the incremental dollar has changed from investment-grade hyperscaler capex to frontier-lab capex, and that is a different security even at the same growth rate. We hold Bear 2 at neutral rather than upgrading it, because breadth improved and counterparty quality did not.
Model Update Needed
The standing model was built on 2026-07-13 against the FQ2 guidance set, and deliberately anchored FY27 AI revenue at management's then-stated "in excess of $100B" floor rather than the Street's extrapolation. That anchoring decision was correct for one quarter and is now stale: management has replaced the floor with a supply-secured number 15% higher and added a year beyond it.
| Item | Current Model | Suggested Change | Reason |
|---|---|---|---|
| FQ3 FY26 actuals | ~$29.35B simulated | $29.591B reported | Replace estimate with filed 8-K figures across IS, BS and CF |
| FY26E AI semiconductor revenue | ~$55.2B | ~$58B | Management raised the framework from ~$56B; FQ4 guided $21.7B |
| FY27E AI semiconductor revenue | ~$103B (floor-anchored) | ~$115B | Floor replaced by a supply-secured number; demand said to exceed it |
| FY27E total revenue | ~$161.0B (7.3% below consensus, deliberate) | ~$170B | AI $115B plus non-AI semis ~$18B plus software ~$37-38B |
| FY28E semiconductor YoY driver | +36% | ~+85% to +90% | FY28 AI of ~$230B against an FY27 semiconductor base near $133B |
| FY28E non-GAAP EPS | derived, no anchor | sanity-check against >$30 | First company-supplied out-year EPS marker; model should reconcile to it |
| Capex % of revenue | 1.2% flat | ~2.5% FY27, ~2% thereafter | FQ3 ran 1.8%, FQ4 guided 4.0%; Singapore substrates plus tripling laser capacity |
| Inventory % of quarterly COGS | 62% | ~50% | FQ3 ran 49.5% against 63.9% in FQ2; supply is tight, not building |
| SBC % of revenue | 7.8% stepping to 5.4% | start at 6.8%, same glide | FQ3 SBC fell 13% in dollars; the ramp is ahead of schedule |
| GAAP effective tax rate | 11% stepping to 12% | ~14% | FQ3 GAAP ETR ran 14.3%; the FY25 net benefit has faded |
| Debt and interest expense | held flat at $62.7B LT | $57.2B LT, less $1.5B post-quarter | $5.6B repaid in FQ3 plus $1.5B of notes retired after quarter end |
| Buyback schedule | $1.0B FQ3, $0.55B FQ4 | $0 FQ3; reassess FQ4 | No repurchases under the program in FQ3 |
| Consolidated gross margin path | cost of products drifting to 22.6% | steeper compression | FQ4 guided to ~73% from 78% a year ago on rising XPU memory content |
Valuation impact: Earnings power moves materially higher on a two-year view while the near-quarter line moves slightly lower. At the $367.24 pre-print close, AVGO trades near 32x our FY26 non-GAAP EPS estimate of approximately $11.61 and roughly 12x the FY28 marker management put on the record. Applying a market multiple of about 22x, the current price discounts approximately $16.70 of FY28 EPS, which implies a 20% two-year earnings CAGR against a company framework calling for roughly 61%. On enterprise value the stock is near 34x an annualised FQ3 free-cash-flow run-rate of $54.7B, against roughly $35.4B of net debt that is actively shrinking. We do not publish a price target, and the framework above is deliberately anchored to the supply-secured FY27 number rather than the FY28 one.
Thesis Scorecard Post-Earnings
| Thesis Point | Status | Notes |
|---|---|---|
| Bull #1: Indispensable merchant supplier of custom AI silicon | Confirmed (strengthened) | AI semis +221% to $16.7B, above the company's own guide; TPU v8i shipping ahead of a later-initiated second source; Jalapeño shipped; MTIA ramps in FQ4; Apple added through 2031 |
| Bull #2: AI networking leadership compounds with cluster scale | Confirmed | Networking +2.5x YoY; Tomahawk 7 at 200 Tbps added; Tomahawk 6 deployed at hyperscalers that buy no Broadcom XPUs; networking now expected to grow as fast as XPUs |
| Bull #3: Extreme operating leverage and FCF conversion | Confirmed (strengthened) | Non-GAAP opex +2.5% on revenue +86% and down 3.9% sequentially; record 67.9% operating margin; 46.2% FCF margin; SBC down 13% in dollars |
| Bull #4: Multi-year visibility de-risks the out-years | Confirmed (extended) | FY27 AI raised to ~$115B and FY28 introduced at ~$230B, both described as supply-secured; first out-year EPS marker at >$30 for FY28. Offset: bookings and backlog no longer quantified |
| Bear #1: Priced for perfection / expectations risk | Challenged | The premium has largely deflated. Stock 23% below the prior print's opening level and 24% below its 52-week closing high, up 6.1% YTD against the index at 12.0%, trading near 12x the FY28 marker |
| Bear #2: Customer concentration in a few frontier labs | Neutral | Breadth improved (six XPU customers plus the Apple ASIC agreement), but the incremental dollar shifted from hyperscaler capex to two private, unprofitable labs that become the two largest XPU customers by 2027 |
| Bear #3: Vendor-financing / circular demand risk | Confirmed (escalated) | ~$29B of maximum exposure confirmed on the first 1 GW tranche, equal to 29% of book equity, against a platform designed for more than 20 GW and no aggregate cap disclosed |
Overall: Thesis strengthened. Three of four bull pillars moved from confirmed to confirmed-and-strengthened on evidence rather than assertion, and the expectations bear point weakened materially as the stock de-rated 23% through a quarter in which the contracted framework expanded twice. Against that, the financing bear point escalated from the watch item it was at FQ2 to the thesis's live risk, on a disclosure refusal rather than on any adverse development. The net is a better business at a better price with one specific, monitorable liability that is larger and less bounded than it was three months ago.
Action: Maintaining Outperform, and we would continue to add on weakness. The position should be underwritten against the supply-secured FY27 framework of ~$115B rather than the FY28 framework of ~$230B, because the market is currently paying for roughly half of the FY28 marker and does not need to be right about it. Watch items for the next print, in priority order: the 10-Q contingent-liability footnote and whether XPV exposure scales without a stated ceiling; whether the FY27 AI number is reiterated or raised again once FQ4 ships; the consolidated gross-margin trajectory below the guided 73%; whether non-AI semiconductors and infrastructure software resume the growth trajectory both missed this quarter; and whether the buyback restarts.