Back to index

The Margin That Isn't There

Part 1: Where the Consensus Came From, and Why It's Incomplete

The market looks at NVIDIA's 75% gross margin and sees pricing power. This is not an unreasonable interpretation. NVIDIA commands 75%+ of the AI accelerator market, CUDA creates genuine switching costs for training workloads, and hyperscalers are spending $600-700 billion annually on AI infrastructure with no signs of slowing. When a company captures three-quarters of a chip's selling price while contributing only 14% of its bill-of-materials, the natural conclusion is that the company has built an unassailable moat.

The consensus formed because the evidence supporting it is real. NVIDIA's data center revenue hit $115 billion in FY2026. The H100 costs roughly $6,400 to produce and sells for $30,000-40,000. Every hyperscaler needs GPUs to stay competitive in the AI race, and no one can afford to be the first to pull back on spending. The CUDA ecosystem represents decades of accumulated software infrastructure that makes switching to alternatives prohibitively expensive for most workloads. These are not imaginary advantages.

But here's what the consensus misses: NVIDIA's margin is not a moat. It's a residual.

A moat is something you control. A residual is what's left over after everyone else takes their cut. NVIDIA's 75% gross margin sits between two sets of counterparties who have independent pricing power and are currently charging below their theoretical maximum. On the supply side, TSMC controls 92% of sub-5nm logic fabrication and 100% of CoWoS advanced packaging at scale. The memory oligopoly—SK Hynix, Samsung, and Micron—controls 95%+ of HBM production, with all capacity sold out through 2027. On the demand side, the four largest hyperscalers represent 40%+ of NVIDIA's revenue and are collectively investing billions into custom silicon alternatives that provide credible bargaining leverage even if they never capture meaningful market share.

The bilateral monopoly structure matters because it determines who captures the surplus as AI infrastructure spending scales. TSMC currently takes roughly 3-4% of a GPU's selling price for fabrication and packaging—a fraction that was set when GPUs were $1,000 gaming products, not $40,000 AI accelerators. The company announced 5-10% price increases across sub-5nm nodes starting January 2026, with CoWoS packaging capacity sold out through 2027 despite quadrupling from 32,000 wafers per month in late 2024 to a targeted 130,000 by late 2026. A 30% across-the-board TSMC price increase would add only $260-280 in marginal cost per GPU—less than 1% of selling price, easily absorbed within NVIDIA's current margin buffer. The arithmetic shows TSMC has structural room to reprice for years before approaching any ceiling implied by the value of its services.

The memory layer compounds the pressure. HBM accounts for 45% of GPU production cost, and that share rises with each generation because the architectural forces driving memory demand compound rather than cycle. Context windows are extending from tens of thousands of tokens to over a million. Sparse mixture-of-experts models load 176 billion parameters to activate only 39 billion per forward pass, but all expert sub-networks must remain resident in memory. NVLink bandwidth doubles each generation, enabling rack-level memory co-scaling. Per-GPU memory has increased 3.6x across three generations: 80 GB for H100, 192 GB for B200, 288 GB projected for Vera Rubin. Samsung and SK Hynix negotiated approximately 20% price increases on HBM3E contracts for 2026. DRAM spot prices nearly tripled year-over-year by Q4 2025. SK Hynix posted 68.8% gross margins in CY2025; Micron's Q2 FY2026 gross margin reached 74.9%, with Q3 guidance at 81%. For the first time in semiconductor history, memory company gross margins are approaching and in some quarters exceeding TSMC's 63-65%.

The customer side creates the other half of the vise. Google's TPU v7 delivers 4,614 peak FP8 TFLOPS with 192 GB HBM, matching NVIDIA's B200 on compute density. Amazon's Trainium2 offers competitive training throughput at roughly one-third the per-chip hourly cost. AMD's MI325X delivers 2,615 FP8 TFLOPS with 256 GB HBM3e. Three years ago, no custom accelerator came within an order of magnitude of NVIDIA's flagship on raw compute. Now that gap has functionally closed, and hyperscaler motivation reduces to margin arithmetic. Cloud providers operating at 50-68% gross margins on compute services are paying a 75%-margin tax to their primary silicon supplier—a margin that exceeds every other major participant in the AI hardware stack. Custom silicon programs are the rational response to a supplier extracting more economic rent than its customers can sustainably absorb.

The critical insight from Nash and Rubinstein's work on bilateral monopoly is that the party with the superior outside option captures a disproportionate share of the cooperative surplus. Three years ago, hyperscalers had no credible alternative to NVIDIA. Today they do, and the threat disciplines pricing even if total ASIC market share plateaus at 20-25% rather than reaching the 33-45% that bulls project. The ASICs don't need to succeed at scale to compress NVIDIA's margin—they need only to exist as a credible alternative that gives hyperscalers bargaining leverage in every procurement negotiation.

What actually happened validates the mechanism. NVIDIA's GAAP gross margin compressed from 75.0% in FY2025 to 71.1% in FY2026—a 390 basis point decline in a single year. Management guided Q1 FY2027 to 73-74%, down another 100-200 basis points year-over-year. CFO Colette Kress acknowledged in the Q3 2026 earnings call that "input costs are on the rise but we are working to hold gross margins in the mid-seventies." The compression is not a future prediction; it's already occurring. The question is whether it continues.

The market is pricing NVIDIA at 37x trailing earnings with the implicit assumption that 75% gross margins represent a floor. But if NVIDIA's margin is a residual between rising supplier costs and contracting customer willingness to pay, then the floor is wherever the squeeze stops—not wherever NVIDIA's pricing power would suggest. TSMC can raise prices 5-10% annually for years without approaching its theoretical maximum. Memory vendors have transitioned from spot-market volatility to long-term contracted supply agreements that sustain 74-81% gross margins. Hyperscalers have custom silicon programs that provide bargaining leverage regardless of whether those programs ever capture significant market share.

I don't know how fast the compression occurs. The mechanism operates gradually, not suddenly, because TSMC's repricing happens through multi-year contract renegotiations and hyperscaler ASIC deployments scale over 18-24 month cycles. The 390 basis point decline from FY2025 to FY2026 could be the start of a multi-year trend toward 68-70% margins, or it could be a one-time adjustment that stabilizes in the mid-70s as management claims. The uncertainty is real.

What I do know is that the market has assigned valuation multiples in roughly the inverse order of competitive exposure. NVIDIA trades at 37x trailing earnings and 46x EV/FCF. TSMC trades at approximately 25x forward P/E—a discount to the S&P 500 Information Technology sector average despite holding 92% share of sub-5nm logic and a packaging monopoly facing 113% demand CAGR. Memory companies trade at 5-12x forward P/E despite 95%+ oligopoly concentration and capacity sold out through 2027. The most competitively exposed layer commands the highest multiple, while the most structurally entrenched layers trade at the deepest discounts.

The positioning question is not "will NVIDIA's revenue decline" but "where does the surplus accrue as AI infrastructure spending scales." Revenue can grow while margin compresses. The chip designer can ship more units at lower profitability per unit. The market has not priced the distinction.

Part 2: The Trade Everyone Sees But Won't Take

The institutional analysis is right about the facts and too cautious about the trade.

Here's what the per-share math actually looks like. NVIDIA's FY2026 data center revenue was $115 billion at 71.1% gross margin. That's $81.8 billion in gross profit. If gross margin compresses to 68% over the next two years—a conservative estimate given TSMC's 5-10% annual price increases and HBM vendors' 20% contract increases—gross profit on the same $115 billion revenue base drops to $78.2 billion. That's $3.6 billion in gross profit evaporation, or roughly $1.45 per share in earnings at NVIDIA's current share count, assuming R&D and operating expenses scale proportionally. At 37x P/E, that's $53 per share of market cap at risk from margin compression alone, before accounting for any revenue growth.

But revenue won't stay flat. Even the bears expect NVIDIA to grow data center revenue 20-30% annually through 2027 as Blackwell ramps and hyperscaler CapEx sustains above $600 billion. Let's say revenue hits $150 billion in FY2028. At 68% gross margin, that's $102 billion in gross profit—still up $20 billion from FY2026 despite the margin compression. NVIDIA's earnings grow, the stock probably goes up, and everyone who worried about margin compression looks wrong in hindsight.

Except that's not the trade.

The trade is NVIDIA versus the suppliers who are repricing. If NVIDIA grows revenue 30% annually but margin compresses 300 basis points, and TSMC grows revenue 25% annually while margin expands 200 basis points, the relative value shifts. TSMC's gross margin guidance is 63-65% for early 2026, up from 62.3% in Q4 2025, with long-term targets of "56% and higher" even after the drag from overseas fabs. The company is guiding margin expansion while NVIDIA is guiding compression. Micron's Q3 FY2026 gross margin guidance is 81%—higher than NVIDIA's—with management stating they expect "68%+ through-cycle margins" as HBM transitions from cyclical spot pricing to long-term contracted supply.

The market is pricing NVIDIA at 37x P/E, TSMC at 25x, and Micron at 19.9x. If you believe the bilateral squeeze thesis, you short NVIDIA and go long TSM+MU in equal weight. The pair trade isolates the margin structure thesis while hedging AI demand exposure—all three companies benefit from hyperscaler CapEx, but only NVIDIA's margin is structurally exposed to compression from both sides.

Let's run the downside scenario to show why the structure survives it. Suppose hyperscaler CapEx decelerates in H2 2026—not a collapse, just a deceleration from 40% YoY growth to 20% YoY growth. NVIDIA's revenue growth slows, the stock sells off 20-30%, and the short side of the pair works. But TSMC and Micron also sell off because they're exposed to the same demand driver. The question is whether they sell off more or less than NVIDIA. If NVIDIA's margin compression thesis is correct, TSMC and Micron have structural tailwinds (repricing power, oligopoly dynamics) that NVIDIA lacks. The pair trade survives the demand shock because the relative value thesis—suppliers capturing more surplus than the chip designer—holds even in a slower-growth environment.

The risk is that NVIDIA's management is right and gross margins stabilize in the mid-70s. If that happens, the short side loses money while the long side (TSM+MU) still works but underperforms NVIDIA. You lose on the pair. But the loss is bounded because you're long the suppliers who benefit from the same AI infrastructure buildout. The worst case is not catastrophic; it's underperformance.

The best case is that margin compression accelerates. TSMC announces 15% CoWoS packaging price increases in 2027. HBM vendors push contract prices up another 25% for 2027-2028 supply. Hyperscalers deploy custom ASICs at 25% of incremental AI CapEx, giving them even more bargaining leverage. NVIDIA's gross margin compresses to 68% by FY2028, the market reprices the stock from 37x P/E to 28x, and the short side delivers 25-30% while the long side delivers 40-50%. The pair returns 30-40% over 24 months.

Here's the part everyone misses: you don't need to be right about the magnitude. You just need to be right about the direction. If NVIDIA's margin compresses 200 basis points instead of 300, the trade still works because TSMC and Micron are repricing in the opposite direction. The bilateral squeeze thesis is not "NVIDIA's margin collapses to 60%"—it's "NVIDIA's margin compresses while suppliers' margins expand, and the market hasn't priced the divergence."

The TL;DR:

The market thinks NVIDIA's 75% margin is a moat. It's a residual. The suppliers are repricing, the customers have alternatives, and the margin is compressing. The only question is whether you believe the arithmetic or the narrative.

Related Research

Journals

Theories

Entities

Validation Results

Factor Context

Appendix: Validation Summary

Hypothesis 1: Bilateral Monopoly Squeeze — HIGH confidence (validated)

Hypothesis 2: CoWoS Packaging Monopoly Repricing — MEDIUM confidence (validated)

Hypothesis 3: Memory Oligopoly Structural Shift — HIGH confidence (validated)

Key Finding: All three hypotheses survive validation. NVIDIA margin compression is occurring (not future prediction), supplier pricing power is confirmed (TSMC +5-10%, HBM +20%), and memory vendors have transitioned to contracted supply with sustained 74-81% margins. Market has not priced the bilateral squeeze mechanism.

Falsification Criteria:

Next Validation: Q2 2026 hyperscaler earnings (MSFT, GOOGL, META, AMZN) for CapEx guidance. If CapEx growth sustains >30% YoY, bilateral squeeze thesis strengthens. If CapEx decelerates <20% YoY, demand shock obscures structural margin compression.