Deep Dive · Broadcom (AVGO) · Sunday, October 25, 2026 · Prices are Friday Oct 23 closes (market closed)
Buy the self-funded half after the Oct 28–29 capex prints. Size the lab-funded half like credit risk.
中文版 / Chinese edition: 博通深度分析
The call: Accumulate AVGO only after the Oct 29 close. Entry $298–325 in thirds, target $367–375 (1–3 months), invalidation a daily close below $285. SPECULATIVE today; HIGH once the Oct 28–29 hyperscaler capex guides hold and the 10Y is below 5.10%.
Why now: AVGO is 35% below its June high and 12.5% below its pre-earnings close ($367.47) after a record quarter. At 16x FY28 EPS, $321.68 prices FY28 AI revenue of about $135B against management's $230B.
The disagreement: The Street (47 Buy, $527.59 average target) argues demand. Bears argue financing. Both skip the arithmetic: about 60% of FY27 AI revenue (~$70B, my estimate) rides on Anthropic and OpenAI. The guarantees cost at most $9 a share; the revenue dependence is worth about $7 of EPS, roughly $115 a share.
The level that changes everything: the 10Y at 5.10% (above it, halve the position; Evercore's alarm is 5.62%). On the stock, $285 is the line.
Nothing on the demand line broke; the price of money and the disclosure of who funds the demand did. On Sept 2 Broadcom reported revenue of $29.59B (+86%), AI semiconductor revenue of $16.7B (+221%, 56% of sales), non-GAAP EPS of $3.32 and free cash flow of $13.7B (46% of revenue) (press release). The Q4 guide was $34.8B against $35.03B consensus, with gross margin guided to about 73% from 75% (CNBC). Long-range guidance went the other way: AI revenue of $115B in FY27 and $230B in FY28, and FY28 EPS "in excess of $30" against LSEG consensus of $25.86.
Since then the stock has lost 12.5% on seven weeks of financing headlines, not demand headlines:
| Date | Event | Read-through |
|---|---|---|
| Wed Sep 2 | Q3 print; Q4 revenue guide $0.23B below consensus | Stock −6% next day to $346.49 (24/7 Wall St.) |
| Thu Oct 1 | Anthropic's IPO prospectus: Broadcom to lend up to $42B via convertible notes, about a third of a $125.2B five-year TPU lease (CNBC/Reuters) | Vendor is now also lender |
| Thu Oct 8 | OpenAI annualized revenue about $50B, not the $68B widely reported (CNBC) | AVGO −4% that day |
| Mon Oct 12 | Hock Tan: "A lot of investors have no clue what they are investing in"; capital "might be the biggest constraint" (Wolf Street) | CEO names the risk himself |
| Oct 14–15 | CoreWeave bond scare; Oracle −~9% (24/7 Wall St.) | Lab/neocloud credit contagion |
| Tue Oct 20 | Evercore: 10Y near 5.62% is where "alarm bells" ring (Seeking Alpha) | Rates ceiling for AI multiples |
| Fri Oct 23 | Close $321.68 | −35.0% from $495 (Jun 3) |
My view: FY27–28 numbers went up while the multiple went down. That is a financing-risk repricing, not a demand repricing, and it can be underwritten with arithmetic.
On my model AVGO trades at 10.7x FY28 EPS, and the only way that is fair is if growth stops. Market cap is $1.536T (4,773.6M shares), net debt $35.4B, enterprise value about $1.57T.
| Year | Non-GAAP EPS | P/E at $321.68 | Basis |
|---|---|---|---|
| FY26E | $11.6 | 27.7x | 9-month actual $7.81 + Q4 from guide $3.80 |
| FY27E | $17.9 | 18.0x | My model, AI revenue $115B |
| FY28E | $30.1 | 10.7x | My model; management "exceed $30" |
| FY28 consensus | $25.86 | 12.4x | LSEG via CNBC |
Run it backwards. At a normal 16x multiple, $321.68 equals FY28 EPS of $20.10, which requires FY28 AI revenue of about $135B: 41% below the $230B guide and only 17% above the $115B FY27 guide. The alternative reading is that the market accepts consensus EPS but applies a 12.4x credit-stress multiple. Either way, the price embeds a stall or a distress multiple, or some blend of both.
The demand ask is also modest. Growing from the $21.7B Q4 AI guide to $115B in FY27 needs about 12% sequential growth per quarter, against +54% in Q3 and +30% guided for Q4. What is stretched is deployment and funding, not orders: Tan said demand "exceeds this outlook" and supply is secured. AVGO is the AI-silicon laggard, −6.1% year to date against Marvell +163.4% (24/7 Wall St.); the market pays up for AI exposure without a vendor-finance book and charges Broadcom for having one.
FY26 AI revenue is mostly paid from the operating cash flow of Google and Meta; FY27–28 revenue is increasingly paid from capital markets. Tan said four of six XPU customers fund themselves and the financing vehicles exist for just two, Anthropic and OpenAI. On the call, Bernstein's Stacy Rasgon counted roughly 10GW for 2027 (about 6GW Anthropic plus OpenAI) and roughly 20GW for 2028 (about 15GW); Tan replied "you can count the number of gigawatts" (Q&A portion of the call). Tan then said Anthropic is on track to be the largest XPU customer in 2027.
My estimate. If revenue per gigawatt is roughly equal across customers (Rasgon's $11–12B), the two labs are about 60% of FY27 AI revenue (~$70B) and about 75% of FY28 (~$170B). Tan did not give the split. At the roughly $40B all-in cost per gigawatt that Rasgon attributed to a competitor, 6.3GW is about $250B of 2027 deployment, against combined lab run-rate revenue of about $115B (Anthropic $65B at end-July, OpenAI about $50B at end-September).
The gap is bridged by three mechanisms, none of which sits cleanly on Broadcom's balance sheet: the XPV platform with Apollo and Blackstone ($35B first tranche, up to 20GW by 2028), a lending facility of up to $42B, and residual value guarantees (RVGs). The CFO declined to update exposure on the call. Here is what the reported numbers do and do not show:
| Check | Finding | Read |
|---|---|---|
| Receivables | $13.7B vs $7.1B at FYE25; DSO ~42 days, down from ~44 in Q2 (derived) | No stretch |
| Inventory | $4.5B (+99% vs FYE25); Q3 build $195M vs $1,366M in Q2 | Build slowing |
| Cash conversion | Operating cash flow $14.2B = 108% of net income; FCF less $2.0B stock comp = $11.6B | Clean |
| Cash taxes | $347M paid in Q3 vs $2,187M provision; 9-month $2.23B vs $3.85B | ~$1.6B YTD cash tailwind will reverse |
| Capital return | Buybacks $0 in Q3 (Q2: $600M); $5.6B debt repaid; $3.1B dividends | My read: dry powder kept |
| Off-balance-sheet | RVG maximum loss $29B on the first $35B tranche (100% default, zero recovery); up to $42B Anthropic facility, notes not expected to be sold before Anthropic's IPO | Uncapped for future tranches |
Anthropic's own prospectus flags "potential conflicts of interest" in Broadcom's dual role, says it deposited cash into a restricted account for Broadcom's benefit in April, and warns that certain defaults could accelerate a substantial portion of its lease obligations (CNBC). Anthropic reportedly lost $42B on $4.6B of 2025 revenue (Reuters, citing a leaked prospectus, as cited by CNBC); it is targeting an IPO at around a $2T valuation.
A total loss on every disclosed guarantee costs about $6 a share; losing the lab revenue costs about $115. Spread the exposures over 4.774B shares:
| Exposure | $B | Per share | % of $321.68 |
|---|---|---|---|
| RVG maximum loss, tranche 1 (100% default, zero recovery) | 29 | $6.07 | 1.9% |
| Anthropic facility, fully drawn and lost | 42 | $8.80 | 2.7% |
| Both together | 71 | $14.87 | 4.6% |
| EPS at risk if lab revenue (~$70B) vanishes | — | $7.14 of EPS ≈ $114 at 16x | 35% |
The EPS line is $70B × 61% segment operating margin × (1 − 16% tax) ÷ 4.94B shares. Bank of America's credit work agrees on the first half: even scaling XPV to 20GW with 100% defaults, maximum RVG loss is about $42B against 2027 free cash flow after dividends of about $85B, and stressed net leverage peaks near 3x (BofA, Aug 10). The credit survives; the equity is hurt through revenue. That is why the bear case is $184, not $300, and why the monitor list is about lab funding, not Broadcom's balance sheet: Anthropic's IPO, OpenAI's roughly $30B round in discussion, XPV tranche two, and CoreWeave/Oracle bond spreads as the read-through.
Macro is a one-way gate on the lab-funded half and nearly irrelevant to the Google and Meta half. Hyperscaler cash flow pays for the latter; capital-market access pays for the former, and that access is priced off these levels.
| Asset class | Now | Regime-changing level | Consequence for AVGO |
|---|---|---|---|
| Rates | 10Y 4.999% (Oct 23); recent peak 5.36% | Above 5.10%: halve. 5.62%: Evercore "alarm bells" | Beta of 1.5; financing cost: SpaceX 10Y IG notes sold at 5.875% yielded 7.20% by Oct 9 |
| Credit | Spreads widening from tight levels | IG CDX above 68.33 (March peak); HY CDX above 411.04 | Evercore: equities did not unglue in 2000, 2018 or 1989 until spreads widened. XPV debt is priced off credit |
| Oil / Fed | WTI $103.27, Brent $116.72; 84% odds of a hike by December (CME, per our Oct 23 daily) | WTI below $92 relieves; hot Oct 29 PCE re-prices hikes | Oil sets the Fed path, the Fed sets the 10Y, the 10Y sets the multiple |
| Physical build | Seaport: only 30GW of 55GW announced for 2027 may get built | Each 1GW slip ≈ $11.5B AVGO revenue ≈ $1.2 EPS ≈ $19/share at 16x | Tan: "somewhat less" than 30GW will deploy over two years |
| Equities | S&P 500 7,623.68; NVDA $210.96 (−13% from $243.37 high) | S&P 7,500 daily close | Below it, systematic selling hits a high-beta name first |
Tan himself puts the boom at 2029–2030 and names capital as the binding constraint. I agree with both: the physical build and demand are intact for Google and Meta; funding is the live variable for the labs.
Nothing before the Oct 29 close, then scale in thirds inside $298–325. The Oct 23 daily set a rule of no new mega-cap AI longs into the Oct 28–29 prints; this call respects it. FOMC is Oct 28, Microsoft and Alphabet report Oct 28 after the close, Amazon and Meta on Oct 29.
| Field | AVGO common |
|---|---|
| Entry zone | $298–325, thirds: first $315–325 after the Oct 29 close if capex guides hold and the 10Y is below 5.10%; second $298–308; third only after the Dec 9 print if the AI guide is intact. Do not chase above $335. |
| Target | $367–375 (reclaims the Sept 2 close of $367.47); stretch $400 (the July pivot) |
| Invalidation | Daily close below $285 (1.7% below the March 30 low of $289.96) |
| Timeframe | 1–3 months for the trade; the thesis is a 12-month value of $378 |
| Conviction | SPECULATIVE now (one signal: valuation). Upgrades to HIGH when a second independent signal agrees: hyperscaler capex holds on Oct 28–29 and the 10Y is below 5.10% |
| Size and payoff | Full size 3% of portfolio. Average cost ~$311 on two-thirds: +19% to $371 against −8% to the stop, about 2.3:1. A 2% position loses ~0.17% of the portfolio at the stop and ~0.46% on a gap to $240 |
Open-call interaction. No directional contradiction. XLE (long, open) is the natural hedge: the $103 oil that lifts it is the same shock that compresses AVGO's multiple. META (open) is an AVGO customer (MTIA), and TACT-CRWV and SPEC-BE-1 load the same AI-financing factor, so the shared rule applies: if META closes below $600 or the 10Y above 5.10%, halve all longs, including this one.
Probability-weighted 12-month value is about $378, 17% above the close, which is why I want a lower entry than today's price. Valued on FY28 EPS, since by October 2027 the market looks at FY28.
| Scenario | Prob. | Measurable trigger | FY28 EPS × P/E | Value | vs $321.68 |
|---|---|---|---|---|---|
| Bull | 25% | Dec 9: FQ1 AI guide ≥ $25B and FY27 $115B reaffirmed or raised; XPV tranche two closes with third-party capital; 10Y below 4.90% | $30 × 19x | $570 | +77% |
| Base | 50% | FQ1 AI guide $23–25B; FY27 AI ≈ $95B (−17% on deployment slip), FY28 ≈ $165B | $23.6 × 16x | $378 | +17% |
| Bear | 25% | FQ1 AI guide below $22B (flat sequentially), or Anthropic IPO pulled, or a financing tranche stalls; FY27 AI ≈ $70B, FY28 ≈ $100B | $15.3 × 12x | $184 | −43% |
Load-bearing assumption: the multiple recovers. At today's ~12x on base-case EPS of $23.6, the stock is worth $283, which is the stop zone. The trade needs the Oct 28–29 and rates triggers to lift the multiple before the Dec 9 print can lift the earnings. If the multiple does not recover, the answer is no trade.
Friday Oct 23, 2026 closes, fetched Sunday Oct 25 17:39 UTC from the financial data handler (the quote timestamps show the fetch time; levels match the Friday close). AVGO close cross-checked against the Investing.com page ("Closed 23/10, 321.68") and TheStreet (Oct 22 close $323.24).
| Symbol | Close | Day | Note |
|---|---|---|---|
| AVGO | $321.68 | −0.48% | Range $320.23–327.70; 52-week $289.96 (Mar 30) – $495.00 (Jun 3) |
| NVDA | $210.96 | −0.59% | 52-week high $243.37 (Oct 6) |
| AMD / MRVL | $516.79 / $228.51 | +1.74% / +2.26% | MRVL trailing P/E 75.9 (Finnhub) vs AVGO 40.1 |
| SMH / SPY / QQQ | $527.20 / $759.97 / $704.40 | +0.74% / −0.52% / −0.01% | |
| GOOGL / TSM / ORCL | $320.21 / $418.51 / $126.16 | −1.24% / +1.49% / +0.49% | GOOGL, MSFT report Oct 28 after close |
| Macro (from our Oct 23 daily) | S&P 500 7,623.68; 10Y 4.999%; WTI $103.27; Brent $116.72 | ||
FY26: 9-month non-GAAP net income $38,631M ÷ 4,945M shares = $7.81. Q4: $34.8B × 66% = $22.97B operating income − $0.70B interest + $0.10B other = $22.37B; 16% tax → $18.79B ÷ 4.94B = $3.80. FY26E $11.6.
FY27: AI $115B + non-AI semis ~$18B ($4.3B a quarter, +5%) + software ~$35B ($8.7B a quarter, "stabilize") = $168B (BofA's FY27 revenue estimate is $168.2B). 64% operating margin = $107.5B; −$2.2B net interest; 16% tax; 4.94B shares → $17.9.
FY28: $230B + $19B + $36B = $285B at 63% = $179.6B; −$2.0B net interest; 16% tax; 4.95B shares → $30.1.
Base: $220B revenue at 64%, FY28 EPS $23.6. Bear: $153B at 60%, EPS $15.3. Probabilities 25/50/25 give $378 (0.25×570 + 0.5×378 + 0.25×184 = $377.5).
Sensitivities: $10B of AI revenue = $1.02 of FY27 EPS (60% incremental margin, 16% tax, 4.94B shares), ~$16 a share at 16x. One point of FY27 gross margin on $168B = $0.29 of EPS. Implied FY28 AI revenue at 16x: $20.10 × 4.95B = $99.5B net income ÷ 0.84 + $2B interest = $120B operating income ÷ 63% = $191B revenue − $55B non-AI = ~$135B.
Where this could be wrong: the load-bearing inputs are FY27 AI delivery and the multiple. If AI revenue is $95B and the multiple stays ~12x, the stock is worth $283. If management's $115B arrives and the multiple recovers to 19x on $30, it is $570. Operating margin is the third input: opex was $2.1B non-GAAP in Q3 (+2.5% on +86% revenue); a 65% gross margin instead of 70% in FY27 costs about $1.4 of EPS.
Rasgon's tally on the call: ~10GW in 2027, ~6 from Anthropic and OpenAI; ~20GW in 2028, ~15 from them. Tan's stated customer plans: Anthropic 1GW (2026), +5GW TPU v8i (2027), line of sight to +10GW (2028); OpenAI 1.3GW Jalapeño (2027), over 5GW (2028); Meta 3GW through 2028. Shares used: 6.3 ÷ 10 = 63% of FY27 (rounded to 60%), 15 ÷ 20 = 75% of FY28. Weakness: equal revenue per gigawatt is an assumption; the memory-rich TPU v8i could carry more content, which would raise the labs' share. Tan said not all 30GW will deploy within two fiscal years.
No open-market insider transactions appear in the 30 days before today; the latest Form 4s are from Sept 17–25. Henry Samueli sold about 0.70M shares at $354–362 on Sept 23 (roughly $250M, about 1% of his holdings) under a Rule 10b5-1 plan adopted Dec 16, 2025, so it is scheduled selling, not a signal (Form 4). Mark Brazeal sold 25,000 shares at $401.33 on July 10. Street: 47 Buy, 3 Hold, 0 Sell; average target $527.59; Tigress $600 (initiated Oct 22), KeyBanc $575, Bernstein $575, Mizuho $530 (TheStreet). 24/7 Wall St.'s model target is $419.38. Bank of America cut its Broadcom credit view to Neutral on Aug 10 (bond spreads +105bp/+118bp on the 2036/2056s, 30–45bp wider than TI and Qualcomm; ratings A3/A−/A−).