Spilled It Weekly · D Waugh's Thesis
Broadcom (AVGO)
$334.59 (Oct 19 close, +2.5% on the day) · 32.4% off the $495 52-week high · ~$1.62T market cap · 42.4x trailing P/E
Grade: A-
Broadcom is the toll booth on the AI highway: custom chips and networking silicon that every hyperscaler buildout pays. Power, not demand, is the bottleneck, which makes the $115B AI ramp a question of when, not if. Buy the expectation-miss selloffs. Watch the FY2027 forecast like a hawk.
1. The business, simply
Two buckets. One builds the AI highway, the other collects software tolls on it.
- Semiconductor Solutions (AI revenue $16.7B, +221%, 56% of total): custom AI chips co-designed with hyperscalers, plus the networking chips that stitch data centers together, plus broadband, storage, and wireless chips.
- Infrastructure Software / VMware ($8.75B, +29%): enterprise software that runs companies' data centers and clouds. Sold by subscription since Broadcom bought VMware for $69B in late 2023.
In simple terms
XPU = Broadcom's name for a custom AI chip built for one customer's exact workload. ASIC = a chip designed to do one specific job, the opposite of a general-purpose chip like Nvidia's GPU that does many jobs. Co-design = Broadcom engineers the chip together with the customer (Google, OpenAI), so the chip fits their data center like a key fits a lock. Networking fabric = the chips and connections that let thousands of AI chips inside a data center work as one machine.
Say it simply: Broadcom is the toll booth on the AI highway. Google and OpenAI design their AI brains with Broadcom, Broadcom builds the custom chips, and its networking chips move the data between them. Then VMware software runs the whole operation. Chips plus sticky software is a combination no pure chipmaker offers.
2. The $16.7B question: is the AI money real?
Why this revenue is different from a chip cycle
AI semiconductor revenue was $16.7B last quarter, up 221% from a year ago and up 54% from just the prior quarter. That is not a cyclical bounce. It is more than half of all company revenue now, and it comes from chips co-designed with specific customers for specific data centers.
The reason it is durable: co-design lock-in. A hyperscaler cannot swap Broadcom out next quarter. Ripping out a custom chip means redesigning the data center around a different architecture, a multi-year project. Google's Ironwood TPU v7 shipped in high volume in the quarter, TPU v8i is in production, and OpenAI's first custom accelerator was delivered. These are not spot-market sales. They are multi-year buildouts with Broadcom's silicon designed into the foundation.
The AI ramp, quarter by quarter
AI semiconductor revenue went from side business to majority of the company in about a year. Figures from Q3 FY2026 earnings, Sep 2, 2026.
Q3 FY2026 actual$16.7B (+221% YoY)
Q4 FY2026 guided (+236% YoY)$21.7B
FY2027 forecast (raised from $100B)$115B
In simple terms
YoY = versus the same quarter last year. QoQ = versus the prior quarter. Growing 54% in one quarter on a $16.7B base is violent growth at this scale. Guidance = management's own forecast for next quarter, the number the stock trades against.
3. Financial health: a leveraged machine
Real debt, covered many times over by cash flow
This is not a fortress balance sheet. Total debt is about $59.4B against $24B in cash, leftover leverage from the $69B VMware acquisition, with debt-to-equity around 0.6 to 0.71. But the machine covers it: trailing twelve-month free cash flow is about $39.4B (44% of revenue), the dividend keeps growing at roughly $2.60 a year, and management just announced a $1.5B accelerated buyback on Sep 28. In September they also issued $4.5B in new senior notes with maturities out to 2056 to redeem nearer-term VMware-era debt. The style is aggressive by design. It works until it doesn't, which is why the debt line is the number to keep watching.
Q3 FY2026: the filed record
Quarter ended Aug 2, 2026. Reported Sep 2. Broadcom's fiscal year ends in early November, so this is their Q3.
Total revenue$29.59B (+86% YoY)
Operating income (+92%)$20.1B
Free cash flow (quarter)$13.7B (46% of revenue)
68%
Non-GAAP operating margin, management's "new normal"
42.9%
Net margin: 43 cents of profit per revenue dollar
25-27%
Return on invested capital, VMware debt included
In simple terms
Non-GAAP = profit with one-time items stripped out, the number management wants you to judge them on. Operating margin = profit from running the business per dollar of revenue, before interest and taxes. Free cash flow = cash left after running the business and building. Accelerated buyback (ASR) = the company pays a bank upfront to buy back its own stock fast, a signal management thinks the stock is cheap.
4. Vs the competition
Three different ways to sell shovels to the AI gold rush
Broadcom does not compete with Nvidia head-on. Nvidia sells the standard general-purpose AI chip. Broadcom sells the custom chip, co-designed for one customer's workload, plus the networking silicon that connects everything. Different seat at the same table.
| Company | AI role | Broadcom's edge |
|---|
| Broadcom | Custom XPUs + networking silicon | 70-80% of high-end Ethernet switching silicon |
| Nvidia | General-purpose GPUs, CUDA software | Hyperscalers co-design with Broadcom to escape Nvidia pricing |
| Marvell | Custom silicon competitor | Broadcom's scale and installed co-design relationships are larger |
| Arista | Network switches (systems) | Broadcom sells Arista the chips inside those switches |
The irony of the model: Broadcom wins when hyperscalers want alternatives to Nvidia, and it also wins selling chips to the switch makers. It collects at two layers of the stack. The networking lever gets less attention than the XPUs but rides the same capex wave: as clusters scale from thousands to hundreds of thousands of chips, networking takes a bigger share of every data center dollar. Dell'Oro sees over $100B in cumulative AI back-end switch spending through 2030.
In simple terms
Merchant silicon = chips sold openly to any equipment maker, versus chips a company keeps for itself. Broadcom's switch chips go into competitors' boxes too, which is why its share looks so large. CUDA = Nvidia's software that locks developers into its chips, the moat Broadcom's custom approach routes around.
5. Valuation: expensive looking back, reasonable looking forward
The whole debate in two multiples
On trailing earnings, 42.4x looks full. On next year's expected profit, about 20.5x forward, with a PEG near 0.35. A PEG under 1.0 says the growth justifies the price. You are paying for AI growth that is already arriving, not hoping for it.
Two ways to price it
Trailing multiple reflects last year's earnings. Forward reflects the AI ramp. Market data verified Oct 19, 2026.
PEG ratio (est.)~0.35
under 1.0 = growth covers price
And the stock sits 32% below its $495 high even as AI revenue tripled. The market already applied a discount. The open question is whether the forward growth, the $115B FY2027 AI forecast, actually shows up. If it does, 20x forward is cheap. If it doesn't, 42x trailing is the warning that was always there.
In simple terms
P/E = dollars paid per dollar of yearly profit. Forward P/E uses next year's estimated profit. PEG = P/E divided by the growth rate. Under 1.0 means you pay less than one multiple point per point of growth, the classic "growth at a reasonable price" signal.
6. The $115B debate
Nearly 7x AI revenue in two years. The entire thesis is this number.
Management raised the FY2027 AI revenue forecast to $115B (from $100B) and sees $230B for FY2028. That is nearly seven times this year's AI revenue within two fiscal years. If it lands, the forward P/E near 20x is a bargain and the stock works. If hyperscaler capex pauses, it is the single biggest expectations cliff in large-cap tech. There is no middle ground at these forecasts.
$58B
FY2026 AI guide (raised from $56B)
$115B
FY2027 AI forecast (raised from $100B)
Then the CEO himself complicated the story. On October 12 in Singapore, Hock Tan said many AI investors "have no clue what they are investing in," compared AI investors to casino gamblers, and said the biggest constraint on AI expansion is capital, not chips. When the CEO of the AI chip toll booth says the risk is financing, not demand, take it seriously. If the AI capex cycle pauses for funding reasons, Broadcom's forecasts are the first casualty.
The counterweight is the power data. Morgan Stanley sees a 57 gigawatt US data center power shortfall for 2026-2028, about a third of total data center electricity demand. Of 12-16 GW of planned 2026 capacity, only about 5 GW is under construction. Transformer lead times run 128-144 weeks. The bottleneck is physics, not customers. Scarcity forces the fastest silicon into the racks that do get power, which favors Broadcom's custom AI chips and networking.
In simple terms
Gigawatt = a billion watts, roughly the output of one large power plant. A 57 GW shortfall means data centers want 57 power plants' worth of electricity that the grid cannot deliver yet. Lead time = how long you wait between ordering equipment and receiving it. At 128-144 weeks, transformers ordered today arrive in 2029.
7. Bull vs bear: toll booth or expectations trap?
The bull case: the best toll booth in AI
- Two AI levers, one capex wave. Custom XPUs plus 70-80% of high-end Ethernet switching silicon. Even if one hyperscaler's chip program slows, networking content per data center keeps rising.
- Power scarcity is bullish. 57 GW shortfall means demand is proven and supply is capped by physics. Every new megawatt converts to revenue, and scarcity favors the fastest silicon.
- Customers are the best companies on earth. Google, Meta, Microsoft, Amazon fund capex from cash flow ($660-690B in 2026 vs $443B in 2025). Microsoft has $80B in unfilled Azure orders. The ROI on their spend is proven.
- Expectation misses are entries. Q3 total-revenue guide of $34.8B vs ~$35.0B consensus dropped the stock 6.5% after hours on a rounding error while AI revenue beat. It recovered on the call. That is sentiment, not fundamentals.
- Refinancing is noise at this growth. 86% revenue growth, 68% operating margins, $39.4B in annual free cash flow against $59.4B of debt. The debt shrinks in significance every quarter.
The bear case: priced for a forecast that might not land
- The $115B forecast is the whole stock. Nearly 7x AI revenue in two years leaves no room for a wobble. A delayed chip, a soft quarter, a cautious guide gets punished at 42x trailing earnings.
- VMware is a melting ice cube. 31-43% virtualization share depending on the measure, down from far higher, with 87% of customers actively shrinking their footprint. It is ballast, not growth. Do not price it as expansion.
- Concentration changes shape with great customers. A few hyperscalers set the co-design terms and can squeeze on price. They move in lockstep, so a capex pause hits all at once. And the ~$42B Anthropic financing backstop is a bet on a private cash-burning company, not on Alphabet.
- Capex can slow even with full funding. Power hookups, 128-week equipment lead times, and the AI monetization gap are real constraints. Money does not fix physics.
- The CEO warned you himself. Hock Tan said the constraint is capital, not chips. If funding tightens, the forecasts crack first.
In simple terms
Backstop = Broadcom promised to cover about $42B of Anthropic's financing package if other lenders don't. If Anthropic thrives, it is free goodwill. If Anthropic stumbles, Broadcom owns the exposure. Status as of Oct 19, 2026: the package was still being syndicated with no close reported (FT, Oct 16). Ballast = the steady weight that keeps a ship stable. VMware's software cash flow steadies the company while the chip business swings big.
8. D Waugh's thesis
In his words, from the thesis conversation
1
The power grid is the bull case. 57 gigawatts of shortfall, only a third of planned capacity under construction, transformers on a three-year wait. Demand is proven, supply is capped by physics. Every new megawatt converts straight to revenue, and scarcity puts the fastest silicon in the racks that get power. That is Broadcom's silicon.
2
When they miss the Street but hit their own numbers, that is the entry. Last quarter the total guide missed consensus by a rounding error and the stock dropped 6.5% after hours on nothing. It recovered on the call. Always check the why first, pushed orders versus real demand softness, but sentiment selloffs on intact fundamentals are the buy.
3
If I am going to be concentrated anywhere, I want it to be with the best companies on the planet. Google, Meta, Microsoft, Amazon are funding $660-690B of capex from cash flow, not hope. Microsoft has $80B in Azure orders it cannot fill because of power. Google Cloud is growing 82% at 35.6% margins. The ROI on their spend is proven, which means the ROI on Broadcom's slice of it is proven.
4
Capex does not slow down because of funding. Dilution and cash flow are available. The real constraints are power hookups and equipment lead times, not money, and those constraints stretch the buildout out rather than killing it. A longer buildout at proven returns is fine by me.
5
Refinancing cost is noise at 86% revenue growth and 68% operating margins. $59B of debt against $39B in annual free cash flow gets smaller every quarter. The debt only becomes the story if growth breaks. So I watch growth, not the debt.
9. What breaks it
Honest risks, ranked
1. The $115B FY2027 forecast cracks (35%)
Nearly 7x AI revenue growth in two years is the single biggest expectations cliff in large-cap tech. A hyperscaler pause, a financing crunch like Hock Tan warned about, or an architecture shift away from Broadcom's co-designs breaks it. At 42x trailing earnings, a forecast cut is not a dip, it is a repricing. Watch the Q4 print in early December: the $21.7B AI guide is the first exam.
2. Customer concentration bites (25%)
A handful of hyperscalers set the co-design terms and can squeeze on price or shift architectures. They move in lockstep, so one capex pause hits all at once. The ~$42B Anthropic financing backstop adds exposure to a private company burning cash to grow. Great customers change the shape of concentration risk, they do not remove it.
3. The funding cycle turns (20%)
Hock Tan said the constraint is capital, not chips. If AI investors pull back, capex pauses regardless of demand, and $150-200B of data center spending is already estimated to be slipping from 2026 into 2027-28. Broadcom's forecasts assume the money keeps flowing. Watch hyperscaler capex guides and credit conditions, not just chip orders.
4. Margin mix dilution (15%)
Custom silicon carries lower gross margins than software. As XPUs grow past 56% of revenue, the blended margin drifts down even while dollars grow. Gross margin is 58.3% now, down 210 bps sequentially. If it keeps sliding while the AI mix rises, the market reprices what the growth is worth.
5. Debt plus regulators (headline risk)
$59.4B of debt against $24B cash is manageable at these cash flows but real, and it only stays comfortable while growth holds. On regulation: a potential China review of networking gear (Broadcom switches are in up to 90% of Chinese state-backed data centers per a September FT report) and EU scrutiny of VMware licensing both hang over the non-AI pillars. Headline risk that compresses the multiple while it plays out.
Kill switch: the FY2027 AI forecast gets cut, a top customer diversifies away from Broadcom's co-designs, or incremental returns on the buildout start falling. Any one of those, the thesis is wrong.
D Waugh's verdict
Grade: A-. The best toll booth in the AI buildout: custom silicon plus dominant networking, funded by customers with proven returns, in a market where power scarcity guarantees demand exceeds supply for years. Buy the expectation-miss selloffs. Three numbers to watch: the $115B FY2027 AI forecast holds, gross margin stops sliding, and hyperscaler capex keeps growing. All three hold, the stock works back toward $495. The forecast cracks, it doesn't.
D Waugh · Spilled It Weekly · October 19, 2026
For education only. Not financial advice. Numbers from Broadcom's Q3 FY2026 earnings release and 8-K (Sep 2, 2026), company filings, and public market data, verified Oct 19, 2026.