Top engineers at Anthropic and OpenAI say AI now writes 100% of their code.

If you're not using AI, you're spending 40 hours doing what they do in 4.

These 100+ Claude Code hacks fix that and help you ship 10x faster.

Sign up for The Code and get:

Claim your free playbook

NVIDIA passed $5T market cap. Most write-ups still frame this as "AI yes" vs. "bubble no." Both are vibes.

The question I want answered: at $208, with FY27 consensus EPS in the $8 range, the multiple at a five-year low, $97B in trailing free cash flow, China priced at zero, and customer balance sheets going the wrong direction — what would have to break for this to be the wrong trade, and how would I see it before consensus did?

That's the work.

Q4 FY26 (quarter ended Jan 25, 2026):

Full year FY26: $215.9B revenue (+65%), $120B net income, $97B free cash flow, $103B operating cash flow.

Margin is expanding into the ramp. From 60.5% GAAP at Q1 FY26 to 75% at Q4, while revenue grew ~55%. Hardware companies catching up to demand normally see compression — capacity costs, mix dilution, supply-chain frictions. NVDA is doing the opposite. That single fact does most of the work in the bull case.

The networking story is what most coverage misses. $31B annualized, +10x in five years, accelerating into Q4. NVLink, Spectrum-X, InfiniBand. If you think of NVDA as a GPU company, you've already missed half the model. They're selling the cluster, not the chip.

Most coverage frames the $97B FCF as a fortress. It's better than that — it's a weapon.

What that buys: capacity prepayments (purchase commitments now extend into CY2027), R&D acceleration on the Rubin/Vera/Feynman roadmap, $100B+ in equity stakes in customers (OpenAI, CoreWeave, Anthropic), buybacks at compressed multiples, and the ability to sit out a downcycle without touching debt markets. AMD and every aspiring competitor has to fund their roadmap from cash flow that's a fraction of NVDA's. NVDA is funding its own dominance.

If hyperscaler capex moderates in 2027, NVDA loses growth. It doesn't lose the ability to run the playbook. The bear case isn't "NVDA goes to zero." It's "NVDA gets re-rated as a cyclical." Real risk. Different risk than people are pricing.

Peer group at TTM:

NVDA is the cheapest growth in the peer group. AMD trades at twice the multiple for half the growth and a tenth of the ROE. Broadcom is at a 40% premium for less than half the growth.

The de-rating people kept waiting for already happened. While everyone debated whether the company was overvalued, earnings ran faster than the stock for twelve months. The bear case at $208 has to argue FY27 will miss by a wide margin, not that the stock is "expensive."

Coverage from ~40 analysts:

Recent revisions skew above consensus: Raymond James raised to $323 (from $291) post-Q4. Average of three most recent: ~$292 (~40% upside).

I weight this less than revisions. Targets are stale by construction. But direction matters: in my framework, consensus PTs lag revision activity by 30-60 days, and breadth is one-sided enough that the next move on PTs is up.

Last 90 days: 32 analysts revised EPS up. 1 down.

The loudest revision signal you'll see on a mega-cap. Strong momentum names typically run 8-3 or 10-5. The 32:1 ratio says the people closest to the model are still climbing, not stepping off.

Revision direction tells you where the next earnings number is moving. With breadth this one-sided, FY27 consensus keeps getting marked up. That's how a stock at $208 still has 30%+ sell-side upside priced in even after the move.

Management's FY27 guide assumes zero China data center compute revenue. Pre-export-control, China ran ~20-25% of DC. Today: limited H200 approvals, no shipments.

Two implications most write-ups botch. First, the bull case works without China — nothing in consensus FY27 relies on a re-opening, and any regulatory easing is unmodeled upside. Second, "priced at zero" means it's no longer a swing factor. People still calling it a "key risk" are pricing what's already priced.

The actual ongoing China risk is competitor development inside China — Huawei Ascend, Biren, Cambricon — taking the domestic market and eventually showing up in third-country deployments. That's a 2027-2028 question, not a 2026 one.

The bull case for NVDA isn't NVDA. It's the four companies cutting NVDA's purchase orders.

2026 capex from MSFT, AMZN, GOOGL, META, ORCL: $660-750B, +67% YoY. Capex/sales:

Twenty years in industrials. HVAC distribution runs 4-6% capex/sales. Heavy industrial in a buildout year hits 12%. Truckload carriers in fleet refresh hit 15-20% and get hard questions on the call. These hyperscalers are spending half their revenue on the buildout. Oracle is spending 86%. There is no precedent in normal capital allocation for this. Closest historical comp: late-90s telecom buildout. Not a flattering reference.

Downstream:

The customers are funding NVDA's growth with debt against an asset base that depreciates ~20%/year. That's the bull case constraint, not its support.

NVDA committed $100B to OpenAI and $10B to Anthropic. OpenAI uses NVDA capital to buy compute from Oracle ($300B contract), AMD ($90B), AWS ($38B), and Microsoft ($250B). Those vendors then buy NVDA chips. NVDA holds equity in CoreWeave, which buys NVDA chips and sells capacity to MSFT and Meta. OpenAI's projected 2026 loss: $14B against revenue of ~$13-15B. Total OpenAI infrastructure commitments through 2035: $1.15T.

Capital is moving in loops. The dotcom vendor-financing comp is overplayed but not absurd. Late January 2026 was a preview: $600B was wiped from NVDA's market cap in one session on a WSJ report that the OpenAI deal had "stalled."

What matters for sizing: the loop has to be sustainable through NVDA's order-book window — call it 2-3 quarters. Right now it is. Inventory is up 8% QoQ and management increased purchase commitments to lock in capacity through CY2027. Bullish demand signal, with execution risk if the curve bends.

The other constraint nobody quantifies in their model: power. Several 1GW+ data center projects are slipping schedules on grid interconnect, substations, and cooling. The capex headline assumes those builds come online. If grid timing pushes Vera Rubin deployments by two quarters, NVDA's CY2026 revenue cap is set by physics, not demand. Worth tracking.

NVDA AI accelerator share: ~80%. AMD: 5-7%. Custom silicon (Google TPU, AWS Trainium, MSFT Maia, Meta MTIA): collectively the bigger threat. TrendForce projects custom ASIC shipments +44.6% in 2026 vs. GPU +16.1%.

The unit gap is real. Dollar gap less so — NVDA captures multiples of revenue per chip. AMD's MI400 just landed at CES 2026 (2nm, 432GB HBM4) but ships in volume against Rubin, not Blackwell. ROCm 7.2 closed the technical gap on CUDA. The deployment gap is wider than ever because the codebases keep growing.

The cadence:

The 10x inference token cost reduction is the under-discussed piece. Inference is the second AI revenue battle (training was the first), and that's where AMD and custom silicon were supposed to compete on memory capacity and price/performance. If Rubin's claim holds, NVDA prices itself back into inference workloads at a level customers can't replicate internally — and lowers the cost of AI use enough to expand total demand instead of saturating it. That's a structural argument, not a hardware one.

In March 2026 the board formally tied executive cash bonuses to FY27 revenue. Jensen's target award: $4M (200% of base salary). CFO Kress and EVPs: $1.5M each (150% of base).

Reads two ways at once. Either management has order-book visibility they're willing to bet personal compensation on, or the incentive structure now biases them toward spending through any revenue deceleration rather than cutting to protect EPS. Probably both. Small structural bias, but it matters if you're modeling FY27.

Current $208 sits between base and bear case fair value. Sell-side consensus at $270 models something closer to the bull case. Not a screaming buy or a clean short. A position that pays you for being right on margin durability and customer capex direction.

Two of these flipping is enough to size down before consensus moves:

None of those are calls about narrative. They're prints on a calendar.

Long, sized for a multi-quarter hold. Methodology says yes on every axis: revenue acceleration, margin expansion into the ramp, 32:1 revision breadth, cheapest growth in the peer group, two-generation product visibility, $97B FCF funding offensive moves, fortress balance sheet (D/E 0.07), record customer backlogs (GOOGL $240B, AMZN $244B), free option on China reopening, management compensation tied to top-line.

The ceiling on this trade isn't NVDA's balance sheet. It's the four customer balance sheets. When MSFT, GOOGL, AMZN, or META starts defending free cash flow ahead of building capacity, this position ends fast — probably faster than the chip cycle that follows.

What ends it isn't competition. It's a CFO at one of those four standing up on a 2027 call and saying "we're moderating capex as we evaluate ROI." That's the line that prices in two days, not two quarters.

Until then: long, watchlist active, six numbers checked quarterly.eets. When MSFT, GOOGL, AMZN, or META starts defending free cash flow ahead of building capacity, this position ends fast — probably faster than the chip cycle that follows.

What ends it isn't competition. It's a CFO at one of those four standing up on a 2027 call and saying "we're moderating capex as we evaluate ROI." That's the line that prices in two days, not two quarters.