Broadcom Versus Micron: Two AI Silicon Models, One Year to Decide

Same bell, same clock, one question: In the AI infrastructure buildout, does the custom chip designer outlast the memory king?

Broadcom versus Micron Technology. Both start at 100 paper points at the close of September 12, 2026. One-year horizon. Total return scoring, including dividends. Broadcom opens as the 60/40 favorite. No substitutions.

That handicap may seem backward given what just happened. Micron is up 224% year-to-date. Broadcom is essentially flat. But the scoreboard tracks price, not business durability — and that distinction is exactly what this match tests.

The frozen card: Broadcom (AVGO) against Micron (MU), two semiconductor companies exposed to the same AI spending cycle, one earning from custom silicon and networking design, the other from high-bandwidth memory manufacturing. The disagreement is simple: which economic model produces the better total return from here, over the next year, when the memory cycle is at its peak and the custom chip business is still accelerating?

The Designer's Play

Broadcom's model is built on relationships, not fabs. The company designs custom AI accelerators for hyperscalers — Google's TPUs, Meta's MTIA chips, OpenAI's Jalapeño accelerator, and Anthropic's compute racks — then sends the blueprints to TSMC for production. For every dollar spent on a custom chip, hyperscalers spend another 40 to 60 cents on Broadcom networking silicon to connect them. The company also pulls in $8.8 billion per quarter from infrastructure software, a stable annuity that has nothing to do with semiconductor cycles.

The numbers from Q3 fiscal 2026: $29.6 billion in revenue, up 86% year over year. AI semiconductor revenue hit $16.7 billion — 56% of the total, up 221% from a year ago. Free cash flow was $13.7 billion, or 46% of revenue. Non-GAAP operating margin sat at 68%.

The backlog is the structural detail that matters. Broadcom holds $73 billion in contracted AI revenue spread over six quarters, backed by multi-year agreements with six major customers. The Google TPU deal extends through 2031. OpenAI signed for 10 gigawatts of accelerators through 2029. CEO Hock Tan guided AI revenue to approximately $115 billion in fiscal 2027 and $230 billion in fiscal 2028.

The risk on the designer side is margin dilution. Q4 guidance projected a non-GAAP operating margin of about 66%, down from the current 68%, because the next wave of custom accelerators requires more memory content — chips that Broadcom sells at lower gross margins than pure networking silicon. There is also the question of concentration: six customers, one foundry in Taiwan, and 18 to 24 month design cycles that leave little margin for error.

The Memory King's Supercycle

Micron is running the most extreme upcycle in memory history. The company reported $41.5 billion in revenue for Q3 fiscal 2026 — a 346% year-over-year surge and more than the entire company made in all of fiscal 2025. Non-GAAP gross margin hit a company record of 84.9%. Non-GAAP EPS was $25.11, up 1,215% from a year ago.

The driver is high-bandwidth memory — HBM — the memory architecture that sits stacked directly beside GPU and accelerator chips, feeding them data at speeds conventional DRAM cannot match. HBM requires more than three times the wafer capacity per bit compared to standard DRAM. Every HBM generation gets more intensive. Supply is fully allocated through 2027. Management said there is no "line of sight" to when supply catches up with demand.

The economics of scarcity are brutal. DRAM prices jumped 60% sequentially in the quarter. HBM4, the newest generation, is ramping twice as fast as its predecessor, with over $1 billion already shipped. Micron guided Q4 revenue to $50 billion, with gross margins around 86%.

But the memory business has a habit of reversing. The industry has survived multiple boom-bust cycles over three decades, each one following the same pattern: pricing power peaks, capacity catches up, prices collapse, margins compress, and the stock sells off. Micron's historic peaks are impressive; its historic troughs are what investors remember.

The Mechanism Board

Two scoreboards, two questions.

The return scoreboard is where price lives. Micron leads by a wide margin. Up roughly 224% this year against Broadcom's flat performance. The stock trades at a trailing P/E of 20.7 compared to Broadcom's 43 — seemingly cheap for a company growing revenue at 167% and reporting 65% operating margins.

The mechanism scoreboard tracks what those numbers are built on.

Margin durability: Broadcom's operating margin is high because custom silicon is differentiated, locked into multi-year contracts, and diversified by software revenue. Micron's 85% gross margin is high because HBM demand exceeds supply — a condition that is powerful but temporary by historical definition. Broadcom wins this variable.

Cyclicality: This is the entire Micron question. The company announced 16 five-year "take-or-pay" supply agreements covering approximately $100 billion in minimum committed revenue, backed by $22 billion in customer cash deposits. Management says these contracts are designed to end the boom-bust cycle, producing margins "well above our peak quarterly margins in any past cycle." The agreements are real, the deposits are real, and the take-or-pay structure is a genuine departure from how memory markets have historically operated. But the contracts run to 2030 — long enough to prove the thesis or reveal its limits. The question for a one-year horizon is whether the agreements prevent a sharp earnings reversal if AI spending slows in the next 12 months. Broadcom, which already operates on multi-year contracts, has a proven answer. Micron has a theory it's testing.

Capital intensity: Broadcom spent $1.25 billion on capital expenditures over the trailing twelve months. Micron spent $25.3 billion — twenty times more. Memory fabs cost $20 billion or more each, and Micron is building facilities in Idaho, Taiwan, Singapore, and New York. That $25 billion in capex against roughly $18 billion in quarterly free cash flow means the company is borrowing against the upcycle to fund capacity that may arrive just as the cycle turns. Broadcom wins this variable decisively.

Revenue diversity: Broadcom earned 30% of Q3 revenue from infrastructure software, completely insulated from semiconductor cycles. Micron earns 100% of its revenue from memory chips. Broadcom wins.

Why the Scoreboard Contradicts the Mechanism

Micron's stock has done what it does in every upcycle: compress the next three years of good news into 12 months of price appreciation. The 224% year-to-date gain implies a future in which AI memory demand stays near-peak, HBM pricing never normalizes, the take-or-pay contracts perfectly smooth cyclicality, and $25 billion in annual capex pays off without a misstep.

Broadcom's stock has done what value traps do: underperform while the market doubts sustainability. The stock fell 5% in extended trading after Q3 earnings, even though revenue beat, AI revenue grew 221%, and free cash flow hit a record $13.7 billion. The drop came because Q4 guidance of $34.8 billion missed the Wall Street consensus of $35.0 billion — a $200 million miss on a company worth $1.65 trillion. That is the kind of sensitivity that only happens when investors have priced in perfection and are waiting for a reason to pull back.

Both reactions are rational. Micron rewards the bull on memory scarcity. Broadcom punishes the skeptic who asks, "can this really grow to $100 billion in AI revenue?" But neither reaction is wrong because neither price captures the full picture.

The Handicap Explained

Broadcom opens as the 60/40 favorite not because it has the better stock chart, but because its economics are structural rather than cyclical. The custom chip business is growing faster than memory — AI revenue at $16.7 billion in one quarter, guided to $21.7 billion next quarter — on an asset-light model that converts 46% of revenue into free cash flow. The software annuity provides ballast when chips slow. The backlog provides visibility. The margin profile, while declining slightly, is earned from differentiation, not scarcity.

Micron gets the sympathy of anyone who has watched a cyclical company generate this much cash in this short a time. The take-or-pay contracts are the most credible attempt to break the memory cycle that the industry has ever seen. If they work — if half of Micron's revenue is truly locked into five-year agreements with price floors — the company deserves a fundamentally higher valuation than the 20x trailing multiple it currently commands. But proving that claim takes years, not quarters. A one-year horizon is the perfect length for a memory cycle to prove it hasn't been tamed.

The odds are not a prediction. They are a reflection of what kind of business each company is, not what either stock has done lately.

The Ledger

Both companies are excellent at what they do. Micron is the world's largest US memory chip maker, producing HBM that NVIDIA, AMD, and every hyperscaler needs. Broadcom designs the silicon that makes AI clusters affordable at scale and connects them together. Neither will disappear.

The match turns on one mechanism: whether Micron's contracts transform its economics from cyclical to contracted, or whether the memory cycle reasserts itself within the 12-month window. Broadcom's economics are already proven. Micron's are being tested.

The scoreboard favors Micron. The mechanism board favors Broadcom. The final bell — one year from now — will tell which economy actually delivered.