Broadcom did not fall 6% on 14 August because AMD took its Google business. That is the story the market settled on, and the timeline does not support it. The SemiAnalysis research note everyone cited as the cause was not reported until 17 August — three sessions after the drop — and on the day it actually landed, Broadcom closed down 0.14%. What moved AVGO on 14 August was a Bank of America credit note: analyst Tom Curcuruto cut Broadcom’s issuer and bond rating from Overweight to Marketweight and put a $370bn number on the off-balance-sheet vehicle financing its AI chips. The stock closed at $380.00 on 18 August 2026, down 3.17% on the session, with pre-market at $379.46 on 19 August (07:40 EDT, stockanalysis.com). Against that spot, this piece sets a $560 bull case and a $265 bear case.
The substitution matters more than the six percent. Markets swapped a balance-sheet story for a competition story, because a competition story is easier to tell: chip A beats chip B, one logo wins, one loses. A residual-value guarantee on a special-purpose vehicle leasing custom accelerators to a single anchor tenant has no logo and no scoreboard. And when you read the SemiAnalysis note on its own terms, the competition framing inverts. Its 2026 cut to Broadcom’s Google TPU volumes is attributed to chip-on-wafer-on-substrate packaging supply, not lost sockets. Its 2027 cut is attributed to Broadcom reallocating capacity to other customers — 55,000 wafers moved to Meta’s MTIA 400 “Iris” programme. That is a company turning demand away because it cannot pack chips fast enough. Priced as share loss, it read bearish. Read correctly, it is a supply story inside a demand boom — and the genuine bear case sits somewhere else entirely, in the financing.
Key facts
- Spot $380.00, close of 18 August 2026, −3.17% on the session; pre-market $379.46 (−0.14%) at 07:40 EDT, 19 August 2026 — stockanalysis.com
- 52-week range $281.87–$495.00 intraday; on closing prices, $289.60 (21 Aug 2025) to $481.57 (2 Jun 2026). Spot is 21.1% below the closing high and 31.2% above the closing low — FinanceFeeds calculation on stockanalysis.com daily closes
- Q2 FY2026 AI semiconductor revenue $10.8bn, +143% year on year, on total revenue of $22,187m — Broadcom Q2 FY2026 results, 3 June 2026
- Q3 FY2026 guidance: ~$29.4bn total revenue, $16.0bn AI semiconductor revenue, reported Wednesday 2 September 2026, after the close
- BofA stress case: up to $370bn of senior obligations at the XPV financing platform by mid-2029 at 20GW scale, including ~$150bn of new issuance in 2027 alone — 24/7 Wall St, 14 August 2026
- Disclosed maximum residual-value exposure of $29bn on the initial ~$35bn transaction, in a 100% default scenario — BigGo Finance summary of the BofA note
- Samsung pact exceeding $200bn through 2030, covering HBM memory, 2nm-and-below foundry and 2.3D/2.5D advanced packaging — CNBC/Reuters, 25 July 2026
- Consensus “Strong Buy”, average target $515–$536 depending on data vendor — $527.88 across 48 analysts polled by S&P Global via stockanalysis.com, $536.10 across 55 via LSEG; median $530, high $675, as at 19 August 2026
- Market capitalisation $1.81trn on 4.76bn shares, trailing P/E 63.2x and forward P/E 24.1x, beta 1.47 — stockanalysis.com, 19 August 2026
What actually happened on the tape
Take the three sessions in order, because the ordering is the whole argument. Broadcom closed 13 August at $417.82. On 14 August it closed at $392.99, down 5.94% — the steepest fall among the thirty constituents of the Philadelphia Semiconductor Index that day. AMD closed the same session at $514.39, up 6.50%. On 17 August, AVGO closed at $392.43, down 0.14%. On 18 August it closed at $380.00, down 3.17%. Cumulatively, $417.82 to $380.00 is a 9.05% drawdown across three sessions.
Now attach the causes. The 14 August move had at least three named catalysts, and not one of them was a TPU. Bank of America’s Curcuruto downgraded Broadcom’s credit, flagging the financing platform. Separately, Baird’s Tristan Gerra doubled his AMD price target to a Street-high $1,250 from $625, maintaining Outperform, on a model projecting AMD AI GPU platform revenue reaching $147bn by 2030. Third, 14 August was the deadline for Q2 13F filings, and Broadcom’s holder list thinned in public: Third Point dissolved its position entirely, Tiger Global cut its stake, and both Stanley Druckenmiller and Dan Loeb exited. Three independent events, none of them a chip. That the move was company-specific rather than sector weather is confirmed by the tape: the iShares Semiconductor ETF fell just 0.7% that session. The market produced a 12.4-percentage-point spread between AVGO and AMD, then wrote a single narrative over the top of unrelated documents.
The SemiAnalysis note is the third document, and it arrives later. Its contents were reported on 17 August: 2026 TPU v7 production revised down from 3.2 million to 2.7 million units, and Broadcom’s total 2026 CoWoS wafer output cut from 250,000 to 215,000 wafers, on CoWoS-S advanced packaging ramp challenges. On the AMD question the note is explicitly hedged. “Market chatter suggests [Google] is working with AMD on a TPU project in the v10 generation,” it says, adding that “AMD’s involvement would be the first real involvement in a custom AI ASIC project” — a tenth-generation part, years from revenue, described as chatter. Covering the same report, Tom’s Hardware framed it as a hybrid design that could integrate on-package CPU cores. Analyst Patrick Moorhead noted publicly that even if the AMD-Google part is real, it could still include Broadcom I/O.
Having tracked how these narratives form across a dozen semiconductor drawdowns this year, the pattern is consistent: when a stock falls on a structural or financial concern that takes four paragraphs to explain, the market will find a competitive explanation that takes one sentence, and adopt it retroactively. It happened to TSMC on packaging constraints and it is happening to Broadcom now. The trading consequence is that the mispricing does not correct when the competitive claim is debunked, because the actual problem was never the competitive claim.
The financing structure that is the real bear case
Broadcom’s AI XPV platform is a joint vehicle with Blackstone and Apollo Global Management, built to finance the deployment of Broadcom’s custom AI accelerators. The initial transaction is roughly $35bn, covering about 1GW of XPU chips. The anchor tenant is Anthropic, on a five-year lease, with OpenAI named as a potential future partner. Broadcom backstops the majority of the debt through a residual-value guarantee, with disclosed maximum loss exposure of $29bn on that first transaction in a total-default scenario.
Curcuruto’s work is a scaling exercise on that structure. If the platform expands to 20GW by 2028 at a 2GW quarterly pace, senior obligations reach approximately $370bn by mid-2029, with around $150bn of new issuance required in 2027 alone. Broadcom’s maximum residual-value guarantee exposure scales with it. BofA’s stress test puts extreme losses — the 100% default case — at roughly $42bn, and models a more realistic 25% default rate at around $10.5bn. The two risks it names are uncertain residual values for XPU chips, because no mature secondary market exists for custom accelerators, and high tenant concentration, with the platform currently relying primarily on a single lessee.
This is a credit call, not an equity call, and the distinction is doing real work. BofA moved the bond rating, not a stock rating. On spreads, the note argues that levels “already reflect, to a considerable degree, market concerns about XPV platform’s credit profile, leaving limited room for further tightening.” An equity market reading that as “Broadcom owes $370bn” has misread it twice over: the obligations sit at the platform, not on Broadcom’s balance sheet, and the $370bn is a maximum guarantee exposure at a scale the platform has not reached.
What makes it a genuine bear input anyway is the residual-value question, and here Broadcom’s own numbers frame the risk. At Q2 FY2026 the company carried $62,655m of long-term debt and $2,252m of short-term debt against $19,628m of cash. Free cash flow was $10,262m in the quarter, 46% of revenue. Broadcom can absorb a $10.5bn loss over a lease cycle without difficulty. It cannot absorb $42bn without a capital event. The bear case is not that the base case happens; it is that the market reprices the option on the tail, and multiple compression does the damage long before any default does. The same mechanism is repricing AI infrastructure credit more broadly, as the move in long-dated Treasury yields this month showed, and as Nvidia demonstrated when it cut its OpenAI Ohio guarantee from $250bn to $105bn.
Anthropic’s role is worth watching independently. It is simultaneously the XPV anchor tenant, a party to large third-party compute leases such as the $9.1bn Riot Platforms arrangement, and has separately explored a Samsung partnership for its own custom silicon. Tenant concentration is a risk that compounds: the same counterparty appears in several places in the AI financing stack at once.
The bull case: what the guidance and the order book actually say
Set the financing aside and the operating business is accelerating hard. Q2 FY2026, the quarter ended 3 May 2026, produced total revenue of $22,187m — $15,009m from Semiconductor Solutions and $7,178m from Infrastructure Software. Adjusted EBITDA was $15,244m, 69% of revenue. Non-GAAP net income was $12,074m. And AI semiconductor revenue was $10.8bn, growing 143% year on year.
“Q2 semiconductor revenue from AI of $10.8 billion grew 143% year-over-year, above our forecast, driven by increasing demand for custom AI accelerators and AI networking,” said Hock Tan, President and CEO of Broadcom, in the Q2 results release. The guidance underneath that comment is the more aggressive number: Q3 FY2026 total revenue of approximately $29.4bn, up around 84% year on year, with AI semiconductor revenue guided to $16.0bn — growth above 200%, and more than half of total quarterly revenue coming from AI silicon for the first time.
The Samsung agreement is the supply-side answer to exactly the constraint SemiAnalysis identified. Announced 25 July 2026, the pact runs five years through 2030 and exceeds $200bn. Samsung supplies HBM and other memory for Broadcom’s future AI accelerators; its foundry arm manufactures Broadcom chips on 2nm and smaller nodes; and critically, the deal covers advanced 2nm-class packaging including 2.3D and 2.5D integration. If Broadcom’s 2026 volumes were cut because CoWoS-S packaging could not ramp, then a second qualified advanced-packaging source is the single most valuable thing the company could have signed. The bear reading of the packaging constraint and the bull reading of the Samsung deal are the same fact viewed from opposite ends of a supply chain.
Here is a synthesis the individual sources do not state. Broadcom’s Q3 AI guide is $16.0bn. Polymarket’s market on the outcome — “Will Broadcom (AVGO) Q3 AI revenue be above __?” — prices 94% on clearing $15bn, 72% on clearing $16bn, but only 32.5% on clearing $17bn and 14.5% on $18bn. Traders therefore think a beat is likely but a large beat is not: roughly a 72% chance of meeting guidance versus a 32.5% chance of exceeding it by more than about 6%. That is a market pricing execution, not upside. One important caveat: this event has traded only about $5,090 of total volume against roughly $1,178 of liquidity. It is thin enough that it should be read as a sentiment sketch, not a probability, and no position should rest on it.
Disclosure, accounting and the tension nobody is regulating yet
The push-pull here is not a regulator versus an innovator. It is disclosure practice running behind financial engineering. Off-balance-sheet vehicles with residual-value guarantees are long-established structures — aircraft leasing has used them for decades — but the underlying asset in aviation has forty years of resale comparables and a liquid global secondary market. A custom AI accelerator designed for one customer’s workload has neither. Residual value is the entire credit question, and it is the one input with no observable market price.
That is why the $370bn figure travelled so fast: it is a large number attached to an asset class nobody can independently mark. Broadcom discloses maximum exposure on the initial transaction — the $29bn figure exists in company disclosure, not merely in a sell-side model. But the disclosure describes one $35bn transaction, while the analytical debate is about a structure that could scale twenty-fold. Investors are being asked to underwrite the scaling path from a disclosure built for the first step.
Two things would resolve the tension, and both are within Broadcom’s control at the 2 September earnings call. The first is naming additional XPV tenants, which directly attacks the concentration risk. The second is disclosing the guarantee mechanics at scale — whether exposure grows linearly with gigawatts, and what triggers a call. Neither requires a regulator. Both require management to decide that a slightly fuller disclosure is cheaper than a persistent valuation discount. The wider market is already applying that discount to AI capex generally, as Morgan Stanley’s warning on Microsoft showed this month.
The bracket: $560 bull versus $265 bear
Both numbers are scenarios, not forecasts, and both are anchored to the verified spot of $380.00 as at the close on 18 August 2026. The bull case is +47.4%. The bear case is −30.3%.
Bull $560. This requires clearing both the $481.57 closing high and the $495.00 intraday high set earlier this year, and it needs three things to hold together. Q3 AI revenue lands at or above the $16bn guide on 2 September and Q4 guidance extends the trajectory. The Samsung packaging capacity converts into shipped units through 2027, relieving the CoWoS-S bottleneck that cost Broadcom volume in 2026. And the XPV concentration question is answered by naming tenants rather than being left to sell-side stress tests. On 4.76 billion shares outstanding, Broadcom’s market capitalisation at spot is $1.81 trillion; $560 implies roughly $2.67 trillion. Priced off the 24.11x forward earnings multiple the market currently applies to the stock, $560 is about 35.5x forward. That is demanding, but it sits well below the Street’s $675 high target and modestly above the $527.88 consensus average.
Bear $265. This deliberately breaks the 52-week intraday low of $281.87, and that needs justifying rather than asserting. It rests on three legs. First, be explicit that this sits below the Street rather than pretend otherwise. One vendor lists a $215.88 low target, but no other data provider corroborates anything beneath roughly $375, so that entry is best treated as an artifact and not as cover. At $265 this scenario is deliberately outside consensus — and that is the point, because no published target models the XPV structure at 20GW. Sell-side models forecast earnings; this bear case is not an earnings case. Second, the mechanism is multiple compression on the financing structure, not an earnings collapse — if XPV scales toward 20GW while tenant concentration stays unresolved, the equity carries an unmarked tail option, and the market reprices unmarked tails brutally. Third, Broadcom has demonstrated the velocity: it fell 25.15% from the 2 June closing high of $481.57 to the 2 July trough of $360.45 in a month, and realised volatility over the past year annualises around 48%, with 22 of the last 60 sessions moving 3% or more. At $265 the market capitalisation would be roughly $1.26 trillion, or about 16.8x forward earnings on the current 24.11x forward multiple — a de-rating, not a distressed price. A further 30% from here is one bad quarter plus one credit headline, not a black swan.
What would invalidate each: the bull dies if Q3 AI revenue misses $16bn or Q4 guidance decelerates, or if XPV issuance in 2027 tracks toward the $150bn stress path without new tenants. The bear dies if Broadcom names two or more additional XPV lessees, or if Q3 AI revenue clears $17bn — the outcome Polymarket currently prices at only 32.5%.
My base expectation is that neither extreme resolves before the fourth quarter. The 2 September print is a guidance event, not a structural one: it can confirm the demand, but it cannot mark the residual value of a custom accelerator, and that is the variable actually setting the multiple. Expect the range between roughly $340 and $440 to hold into the print, with direction decided by tenant disclosure rather than by the revenue line. For the competitive comparison the market got wrong on 14 August, our AMD bracket of $770 bull versus $350 bear covers the other side of that trade, and the optical transceiver supply chain carries much of the same networking demand.
Frequently asked questions
Why did Broadcom stock fall on 14 August 2026?
Bank of America analyst Tom Curcuruto downgraded Broadcom’s issuer and bond rating from Overweight to Marketweight, flagging that the company’s XPV AI chip financing platform could carry up to $370bn of senior obligations by mid-2029 at 20GW scale. AVGO closed down 5.94% at $392.99. The widely repeated claim that AMD had taken Broadcom’s Google TPU business was not the cause — that research note was not reported until 17 August.
Did AMD really take Broadcom’s Google TPU business?
Not on the evidence available. The SemiAnalysis note reported on 17 August described AMD’s possible involvement in a tenth-generation Google TPU as “market chatter”. Its 2026 cut to Broadcom volumes was attributed to CoWoS-S packaging supply constraints, and its 2027 cut to Broadcom reallocating capacity to other customers, including 55,000 wafers to Meta’s MTIA 400 programme — reallocation, not displacement.
What is Broadcom’s XPV platform?
XPV is a joint AI chip financing vehicle with Blackstone and Apollo Global Management. The initial transaction is around $35bn covering roughly 1GW of XPU accelerators, with Anthropic as anchor tenant on a five-year lease. Broadcom backstops most of the debt via a residual-value guarantee, with disclosed maximum exposure of $29bn on that first transaction in a total-default scenario.
When does Broadcom report Q3 FY2026 earnings?
Wednesday 2 September 2026, after the US market close, with the conference call at 2:00 p.m. Pacific / 5:00 p.m. Eastern. Guidance is approximately $29.4bn of total revenue and $16.0bn of AI semiconductor revenue, which would represent more than 200% year-on-year growth in the AI segment.
What is the analyst consensus on AVGO?
As at 19 August 2026, 48 analysts polled by S&P Global rate Broadcom a consensus “Strong Buy” with an average price target of $527.88, a median of $530, a high of $675 and a low of $215.88. Spot of $380.00 sits 38.9% below the average target and 76% above the lowest.
Is the $265 bear case realistic given the 52-week low is $281.87?
It breaks the range low deliberately, and it sits below every corroborated Street target — those bottom out near $375. That is intentional rather than accidental: sell-side models forecast earnings, while the bear mechanism here is multiple compression on an unmarked residual-value exposure rather than an earnings miss. Broadcom fell 25.15% in the month to 2 July 2026 and realised volatility annualises near 48%, so a 30% decline is within demonstrated range.
Spot price of $380.00 is the close on 18 August 2026, with pre-market of $379.46 verified at 07:40 EDT on 19 August 2026 via stockanalysis.com. The $560 bull and $265 bear figures are FinanceFeeds scenarios for analysis, not price targets, recommendations or investment advice. Bracket assertion: $265 < $380.00 < $560. Do your own research.