Nvidia: Sell Before It’s Too Late

byTony Hicks -
0


This is opinion and not financial advice. Do your own research.

TL;DR

  • Nvidia just printed astonishing numbers—but they’re already yesterday’s story. When growth shifts from frenzied AI training to everyday inference, hyperscalers have every incentive to cut costs with custom silicon and cheaper alternatives. Google (TPUs), Amazon (Trainium), Microsoft (Maia), and Meta (MTIA) are all moving there. Google Cloud+1Microsoft AzureReuters

  • Power and grid constraints—not desire—are the gating factor on new “AI factories.” Capex can only go as fast as electricity, transformers, and land allow—slowing deployments and elongating refresh cycles. MicrosoftCGEPThe Economist

  • The scarcity premium on Nvidia GPUs is fading as supply rises and clouds slash prices (AWS: up to 45% cuts on GPU instances). Scarcity rents don’t last; margins normalize. Amazon Web Services, Inc.

  • Supply-chain dependence is a single point of failure: TSMC’s advanced packaging (CoWoS) remains the bottleneck, with Nvidia historically consuming a massive share of capacity. Any wobble there shows up directly in shipments and mix. ScienceDirect

  • Valuation still assumes a long runway of near-perfect execution. Even bulls admit it’s rich by historic semiconductor standards; one widely tracked data source recently had forward P/E around ~40x (and commentary elsewhere frames it even higher). Multiple compression is an ever-present risk. ValueInvestingBarron's

  • Geopolitics and regulators add real downside tails: China export whiplash has already forced charges; U.S./EU authorities are explicitly probing Nvidia’s market power. NVIDIA NewsroomReuters+1

If you’ve enjoyed a historic run, this is where you ring the register.

1) The setup: an empire at peak velocity

Start with the facts. In late May, Nvidia posted $44.1B in quarterly revenue, up 69% year-over-year, with $39.1B from data center alone. Immaculate? Not quite. That same release disclosed a U.S. license requirement that effectively froze shipments of the H20 China variant and led to a $4.5B inventory/purchase-obligation charge. GAAP gross margin fell to 60.5%; excluding the charge, non-GAAP GM would have been 71.3%. Stellar numbers—and a reminder that the growth engine is not bulletproof. NVIDIA Newsroom+1

Wall Street can love a story until the spreadsheet wins. At the current starting point—hypergrowth, sky-high margins, and a dominant share—the asymmetry shifts. There are more ways to disappoint than to surprise.

2) The capex mirage: it’s not “unlimited,” it’s “electrified”

Yes, hyperscalers are spending like it’s the space race. Alphabet lifted 2025 capex guidance to about $85B; Meta’s range is $66–$72B, explicitly tied to AI infrastructure. Microsoft has talked openly about gigawatt-scale expansion—~2 GW in a single market—and indicated quarterly capex pacing near $30B. Those are staggering numbers. But they are also bounded by physics and permitting. Data centers are increasingly power-constrained, not desire-constrained. MicrosoftGeekWire

Energy researchers estimate U.S. AI data centers alone could need ~14 GW of new power capacity by 2030. The Economist and industry trackers have been blunt about bottlenecks—from high-voltage equipment to grid interconnects. The AI buildout is real; the tempo is the question. When deployments slip for power, not budget, orders shift right—and the scarcity narrative breaks. CGEPThe Economist

3) From scarcity to price wars: the margin trap

Nvidia’s last 18 months were defined by supply scarcity—and customers paying almost anything for Hopper-class GPUs. Scarcity premiums don’t last. As supply improves and second-sources emerge, prices are the release valve.

Case in point: AWS cut prices up to 45% on several Nvidia GPU instance families this summer. That’s not theoretical; it’s live price pressure from your biggest channel partner. Meanwhile, independent trackers show H100/H200 cloud hourly rates sliding throughout 2025 as availability improves. As the market rebalances, the unit economics drift toward normal, not magic. Amazon Web Services, Inc.Jarvis Labs Docs

Gross margin staying at 70%+ in the face of falling per-GPU rents, power-driven deployment delays, and intensifying competition? That’s a high-wire act.

4) The inference pivot: where Nvidia’s moat is shallowest

Training is a sugar high: huge clusters, headline FLOPS, breathtaking pressers. Inference is the diet: relentless, everyday workloads where TCO wins. And the giants are designing silicon precisely for that world.

  • Google runs inference and training on TPUs (v5p, Trillium; and now “Ironwood” for inference), with pod-scale fabrics that bypass the GPU tax. Google Cloud+1blog.google

  • Amazon has Trainium2 in GA and keeps signaling a push to diversify away from “a single chip provider” for cost reasons. About AmazonBusiness Insider

  • Microsoft’s Maia 100 is shipping internally; Redmond keeps telling you it will do more silicon, not less. Microsoft Azure

  • Meta’s MTIA (2nd gen) is live at scale for recommender inference—the biggest inference footprint on earth. GPUs aren’t vanishing, but MTIA’s North Star is lower TCO vs GPUs on its dominant workloads. aisystemcodesign.github.io

Add AMD. MI300/MI325X are real, software has caught up (PyTorch on ROCm; Triton kernels targeting AMD), and MI350/MI400 are on deck. The CUDA lock-in story isn’t gone—but it’s not immovable anymore. Every notch of portability erodes pricing power. AMDROCm DocumentationGitHub

Analysts across the spectrum expect inference to dominate compute needs as apps scale. That’s exactly where custom ASICs and cost-optimized accelerators are most dangerous to Nvidia’s premium. Financial Times

5) China: geopolitics doesn’t care about your product roadmap

Export rules already forced Nvidia into a painful pivot. In April, the U.S. required licenses for H20 exports to China; Nvidia disclosed a multi-billion-dollar charge tied to that whiplash. Reports since then describe a strategy of lower-spec chips for China (e.g., a lower-cost Blackwell variant using GDDR7, not HBM), but this is a margin-dilutive and strategically constrained path—and prone to further policy shocks. NVIDIA NewsroomReuters

Could Beijing or Washington change the rules again with zero notice? Of course. Your position size shouldn’t assume otherwise.

6) Fragile supremacy: CoWoS, HBM, and single points of failure

Nvidia’s brilliance is as much logistics as silicon. But the supply chain is still a Rube Goldberg machine whose most complex gears live at TSMC and in HBM memory.

A 2025 peer-reviewed study quantified what many suspected: Nvidia consumed an outsized share of TSMC’s advanced packaging (CoWoS) capacity—on the order of ~44–48% at points in the cycle. That’s leverage—and exposure. Any hiccup in CoWoS or substrate supply, and entire product waves can slip. ScienceDirect

Yes, packaging capacity has expanded dramatically, and Nvidia has booked capacity aggressively. But concentration risk isn’t a moat; it’s a key-man risk with fab tools. In bull markets, it magnifies profits. In sideways markets, it magnifies volatility. Tom's Hardware

7) Regulators have entered the chat

When you’re this dominant, you attract company. The U.S. Department of Justice is probing AI chips and potential anticompetitive practices. French authorities raided Nvidia’s local offices last year in a related inquiry. Outcomes are unknown, but remedies—from unbundling software/hardware to curbs on preferential supply—would crimp strategy flexibility and, potentially, pricing. Reuters+1

Regulatory overhangs rarely reward elevated multiples.

8) Valuation: perfection with a side of cyclicality

Depending on which source and cut you prefer, Nvidia still trades at hefty forward multiples for a hardware-first business. One widely used tracker recently pegged forward P/E in the high-30s; other market commentary places it higher. In normal semi cycles, even champions don’t wear that crown for long. ValueInvestingBarron's

Two things can bring a titan to earth:

  1. Rate of change slows—inevitable as comparisons harden and deployment bottlenecks bite.

  2. Multiple compression—investors decide the future is still bright, just less miraculous.

When both happen together, the drawdown is not a comment on greatness; it’s a comment on math.

9) “But the demand!”—the bull case, steel-manned

It’s strong, and it’s rational:

  • AI workloads are compounding. AI “agents,” multimodal, and RAG are moving from demos to durable revenue.

  • Hyperscalers’ public capex numbers are the closest thing to a floor you’ll see in tech. Alphabet at $85B; Meta at $66–$72B; Microsoft signaling $30B in a single quarter—the commitment is breathtaking. GeekWire

  • Nvidia continues to innovate at the system level (NVLink/NVSwitch, Grace CPUs, rack-scale NVL72, and 800V DC power initiatives that anticipate the next bottleneck). Platform depth matters. NVIDIA NewsroomNVIDIA Developer

All true. The bear argument isn’t “demand disappears.” It’s that pricing power and mix evolve in ways the current multiple doesn’t fully respect; that the tempo of deployments is bounded by power and permitting; and that the buyers—a tiny oligopoly—are highly motivated to lower their Nvidia bill.

10) The inference economics that quietly erode premiums

Hard numbers keep piling up that inference costs are falling fast—280x cheaper for GPT-3.5-class performance between late 2022 and late 2024, per Stanford’s AI Index. As efficiency rises, the market pushes spend to where TCO per token is lowest—often not on a top-shelf, general-purpose GPU cluster. That’s exactly the wedge for custom silicon and for price competition between accelerators. Hai ProductionStanford HAI

Hyperscalers won’t voluntarily keep paying peak scarcity prices when the cost curve is collapsing and alternatives are good-enough. Amazon’s own leadership has been explicit about driving cheaper AI via chip improvements and supplier diversification. Business Insider

11) Depreciation and the hidden P&L gravity

Rapid product cycles (Hopper → Blackwell → Rubin) thrill engineers and torment CFOs. Shorter useful lives mean higher depreciation and a larger non-cash drag on operating income—precisely what some large customers have already signaled as they update useful lives on AI gear. If your core buyers are re-underwriting depreciation schedules, they’ll pace purchases more carefully—or demand lower prices to compensate. Business Insider

12) What the next 12–24 months could look like (the sober version)

  • Demand stays strong, but sequenced by power/land; mega-campus timelines slip from quarters to years. The Economist

  • Pricing continues to normalize as supply catches demand and clouds use price strategically to win customers. Amazon Web Services, Inc.

  • Mix tilts marginally away from the highest-ASP, highest-margin configs toward cost-optimized racks and more inference-oriented fleets (including non-Nvidia silicon). Google CloudMicrosoft Azureaisystemcodesign.github.io

  • Regulatory noise persists, periodically denting sentiment and complicating bundling/partner strategy. Reuters+1

  • China remains volatile—policy-dependent variants, lower specs, and headline risk baked in. Reuters

That’s not a disaster scenario. It’s just not the “up and to the right forever” scenario embedded in a premium hardware multiple.

13) “Okay, so why sell?”

Because great companies at peak narrative with peak margins and peak customer concentration usually transition to great companies at normal valuations—the path there is called drawdown.

If you bought early, you’ve been paid for your foresight. If you bought later, you’ve been paid for your courage. At this stage, you’re mostly being paid for ignoring obvious mean-reversion forces:

  • Buyer power is consolidating.

  • Alternatives are rising.

  • The grid—not the TAM slide—sets the pace.

  • Regulators are circling.

  • Prices are falling at the edge.

Nothing in that list requires a crash. It merely requires math.

14) “What could prove me wrong?”

Three things:

  1. Power unlocks much faster than expected (e.g., nuclear/behind-the-meter solutions go mainstream by mid-decade), letting hyperscalers pull forward multi-gigawatt campuses.

  2. Software gravity (CUDA + ecosystem) remains so strong that even inference prefers Nvidia for years, and competitors’ roadmaps slip.

  3. Rubin arrives exactly on time, with meaningful perf/watt and cluster-scale advantages that keep ASPs and margins elevated.

All are possible. None are free options.

15) A practical playbook (if you disagree with “sell”)

If you insist on staying long, at least de-risk the path:

  • Treat Nvidia as a cyclical platform leader, not a utility. Size accordingly.

  • Watch the price curve (cloud GPU hourly rates), hyperscaler capex detail (especially power-constrained markets), and regulatory headlines. When price cuts outpace unit growth, margins blink. Amazon Web Services, Inc.

  • Track non-Nvidia silicon adoption signals: public preview → internal GA → external GA timelines for TPU, Trainium, Maia, MTIA, and AMD MI-series. Each step shifts bargaining power. Google CloudAbout AmazonMicrosoft Azureaisystemcodesign.github.io

16) The bottom line

Nvidia built the picks and shovels for the first phase of the AI rush and deserved every ounce of credit—and every dollar of shareholder return. But markets pay for change at the margin, not for greatest-hits reels.

  • The buyers now know exactly what they need.

  • The grid will ration who gets it, and when.

  • The alternatives are good enough to force price discipline.

  • The referees (regulators) have shown up.

  • And the multiple is still priced for magic.

You don’t need to predict doom to justify a sale. You just need to recognize that the reasons you bought have matured into reasons to manage risk. In semis, perfection is a temporary state; gravity is permanent.

Sell before it’s too late.

Sources

Again: this essay expresses a viewpoint. It’s your money and your call.

Tags

You may like these posts

Post a Comment

Previous Post Next Post