NVIDIA's Second Leg: The Compute Selloff Overlooks the Networking Business

A curious thing happened to the most-owned stock in the world this spring: it got cheap. Since peaking in mid-May, NVIDIA shed roughly 16% of its value — about a trillion dollars of market capitalization — even as its forward revenue estimates kept climbing, leaving the shares at $210.96 as of the July 10 close and inside a 52-week range of $162.02 to $236.54 . The proximate cause is not a stumble in the business but a rotation within the AI trade itself: as the acute GPU shortage of 2024–25 eased, the spot rental price of an H100 rolled over from a May peak, while DRAM prices spiked and memory names like Micron nearly tripled, pulling capital toward the new perceived bottleneck .

The narrative that congealed is unusually literal — NVIDIA's equity value is now spoken of as a bet on the price of compute, and that price is falling. Layered on top was a scare from research shop SemiAnalysis that NVIDIA's next-generation Kyber NVL144/NVL576 rack, the housing for the Rubin Ultra generation, could slip to 2028 on a PCB-midplane manufacturing failure — a report NVIDIA has publicly disputed as noise .

That is the setup the market has priced. The argument here is that it misreads what NVIDIA has become. The company is no longer a merchant-GPU vendor whose fortunes track a spot rental rate; it sells rack-scale AI systems in which the network is an increasingly large share of the bill of materials, and that second leg is compounding faster than the compute it wraps. Three things follow, and each is demonstrable from the last filing rather than projected: the networking business is now a genuine second engine the sell-side still carries as a rounding item; the demand in the guide is broad and explicitly excludes China, so the trajectory does not lean on the very compute-price dynamic the bears are fixated on; and at roughly ~16–18x FY28E EPS, the stock has de-rated to a level that no longer demands heroics of the business to work. The one honest fault line — that the FY28 acceleration hinges on a manufacturing cadence a credible shop just questioned — is real, dated, and resolves at the next print, and I treat it as the swing factor rather than paper over it.

Two segments, one company: place the whole business first

NVIDIA reports in two segments, and the imbalance between them is the entire story. Compute & Networking — the Data Center accelerators, InfiniBand, Spectrum-X Ethernet, NVLink, plus the automotive and autonomous-vehicle platforms — generated $74.6 billion of the $81.6 billion total in Q1 FY27 (the quarter ended April 26, 2026, reported May 20), up 88% year-over-year; Graphics, which houses GeForce gaming GPUs and RTX workstation cards, contributed $7.1 billion, up 58% . Operating income tells the same tale more starkly: Compute & Networking threw off $53.3 billion of segment operating profit against Graphics' $2.9 billion . In the older product-line disclosure the shape is familiar — Data Center is roughly 90% of revenue, with Gaming, Professional Visualization, and Automotive rounding out the rest . Automotive is small but strategically live as the autonomous-driving cycle matures; Gaming and ProViz are healthy, Blackwell-driven, and beside the point for the thesis. The company's own re-segmentation this year — splitting Data Center into Hyperscale and "ACIE" (AI Clouds, Industrial & Enterprise), and folding gaming/ProViz/auto into an Edge Computing platform — is management telling you where to look. So will I.

The networking leg the compute-only models miss

The number that reframes the selloff sits one line below the headline. Data Center networking revenue reached roughly $15 billion in Q1 FY27 and nearly tripled year-over-year, while Data Center compute grew 77% to about $60 billion . Put those together and networking has climbed to about a fifth of Data Center (≈$15B of the $75B segment) from roughly an eighth a year earlier — an attach rate that rises structurally as the unit of sale shifts from a loose GPU to a fully-cabled NVL72 rack. Underneath it, InfiniBand grew more than fourfold year-over-year on next-generation XDR deployments, and Spectrum-X, NVIDIA's purpose-built Ethernet fabric for AI, is now by the company's account larger than all Ethernet peers combined . This is the mechanism a compute-centric model cannot see: when a hyperscaler buys a rack rather than a chip, the switches, cables, and interconnect attach at a rising dollar ratio per accelerator, and that ratio is expanding, not holding flat.

Why does this defang the "victim of the compute marketplace" thesis? Because the H100 rental-rate datapoint bears cite should be treated carefully: it reflects pricing for prior-generation Hopper capacity in the cloud rental market, not pricing for NVIDIA’s current Blackwell systems. NVIDIA’s own July pricing slide showed H100 AI Cloud GPU pricing up 13% over three months and 21% over six months, while management says Blackwell now accounts for the majority of Data Center shipments/revenue . The prices NVIDIA realizes selling integrated Blackwell systems where the network, the NVLink switch fabric, and the CUDA software stack travel together. You cannot buy a competing accelerator and get Spectrum-X and NVLink for free; the integration layer is the moat, and it is precisely the part of the P&L growing fastest. That is also why the custom-ASIC threat — the Meta and Google in-house silicon programs that reliably rattle the stock, most recently on reports Meta will begin manufacturing its own AI chip — is shallower than it reads. An ASIC can contest a compute socket. It does not, on its own, disintermediate the rack-scale network that increasingly carries the economics.

A raise built on demand, not price — and with China set to zero

The quality of the quarter matters as much as the size. Q1 FY27 revenue of $81.6 billion rose 85% year-over-year, and management guided Q2 to $91 billion ±2%, with sequential growth expected to come primarily from Data Center . The critical qualifier: that guide assumes no China Data Center compute revenue at all — the U.S. government has approved H200 licenses for China, but NVIDIA has recognized nothing and remains uncertain whether imports will be allowed, so consistent with the prior quarter it embeds zero . The raise is therefore entirely non-China demand. Any policy thaw — and there is live reporting that Beijing may permit limited H200 purchases by top AI firms — is upside sitting outside the numbers, not a plug holding them up.

Within Data Center, the composition is the tell on durability. Hyperscale revenue of about $38 billion was roughly half the segment and grew 12% sequentially; ACIE — the non-hyperscaler half — was about $37 billion and grew 31% sequentially, with AI-cloud revenue more than tripling year-over-year . Demand broadening beyond the concentrated hyperscaler base at more than twice the sequential pace is the early evidence against the "handful of buyers whose capex could blink" fear. It doesn't erase the concentration — three direct customers were 21%, 17%, and 16% of total revenue in the quarter, and one unnamed AI lab is a meaningful indirect buyer through cloud partners — but the fastest-growing dollars are coming from the more diversified side.

Margins held through all of this. Non-GAAP gross margin was 75.0%, essentially flat sequentially through the Blackwell mix, and management guided Q2 to 75.0% ±50bp with the full year in the mid-70s . The year-ago comparison flatters the trend — that quarter carried a $4.5 billion H20 excess-inventory charge tied to the China export ban — but the absolute 75% through a full architecture transition is the durable signal, not the delta.

The cash lens is the honest one, and it reframes the multiple

The reflexive screen on a semiconductor is EV/EBITDA, and here it actively misleads, because it treats NVIDIA as an asset-light 75%-margin compounder and misses where the capital actually sits. Reported capex was a trivial $1.76 billion in the quarter . The real capital base is off the balance sheet: $119 billion of manufacturing, supply, and capacity commitments as of April 26, of which $95 billion falls due within FY27 . That is the money at risk, and it flows through working capital into operating cash — which is why free cash flow, not EBITDA, describes the economics. Operating cash flow was $50.3 billion in the single quarter, against that sub-$2 billion of capex . The incremental-return question is not about capex intensity; it is whether that $119 billion supply book converts to Blackwell and Rubin revenue at the ~75% gross margins the model already earns. At those conversion economics the spread over a ~12% cost of capital is enormous, and the market is characteristically slow to credit off-balance-sheet capital that never shows up in a capex line.

On that lens, the de-rating is the opportunity. NVIDIA now trades around 22–24x consensus forward / NTM earnings — near the S&P 500’s ~20.5x forward multiple and roughly in line with the Nasdaq 100 — despite revenue growing 85% year over year in fiscal 1Q27 and Data Center revenue growing 92% . For a franchise whose networking leg is tripling, whose margins are holding through a node transition, and whose guide excludes its second-largest historical market, that multiple is not pricing continued dominance; it is pricing doubt about the cadence. Tenzing's own build lands FY27 revenue near $393 billion — essentially the consensus average — but reaches roughly $558 billion in FY28 on the networking attach and Rubin ramp the compute-only frames underweight, versus a Street band that spans $391 billion to $643 billion precisely because it has not settled the networking question. On a disciplined ~26x EV/forward-FCF exit that resolves to a base-case value of $265 over twelve months, about 26% above the current price, with a DCF at a 12% cost of equity cross-checking to $283 .

What actually breaks it: the Rubin cadence, not the compute price

The bear case worth taking seriously is not the memory rotation — that is a flow story, not a fundamental one, and the business case is intact — nor the ASIC threat, which is thoroughly in the price. It is the asymmetry of that $119 billion commitment book against a next-generation ramp that a credible research shop has publicly questioned. The FY28 step to ~$558 billion assumes the Rubin platform ships on its second-half-FY27 schedule and converts the supply book on time. SemiAnalysis's report that the Kyber NVL144/NVL576 rack slips to 2028 on a PCB-midplane failure strikes at exactly that assumption; NVIDIA disputes it, and the current Rubin systems are still described as shipping in 2027 to its major cloud partners, but the report is unreconciled and it is the one overhang a refutation alone cannot close .

The reason this matters more than a typical roadmap rumor is the shape of the commitments. The supply agreements are in certain instances cancellable or reschedulable before firm orders, but a meaningful portion is not, and the H20 episode is the disproof-in-miniature: when a single demand pillar was pulled by force outside NVIDIA's control, $4.5 billion converted to an inventory charge in one quarter . A Rubin slip of two-to-three quarters would push the mechanism behind the FY28 acceleration to the right, compress that year toward the consensus low near $391 billion, and re-rate the multiple as the annual-cadence moat narrative cracks — both legs firing at once, which is why a serious bear case lands near $178, roughly 15% below today, not a shallow trim . That resolves at the August 26 print, where the tell is not the headline but whether the commitment book gets marked down or the roadmap re-based.

I discount it, but not to zero. The refutation is a company denial against a third-party note, and denials do not close forensic questions — a disclosed number does. What tilts me is that the operating tape is pointed the right way into that print: a raise that embeds zero China, networking tripling, ACIE broadening the base, margins holding. The capital-return posture buffers the wait — the board added an $80 billion buyback authorization on May 18 on top of $38.5 billion remaining, the company repurchased 108 million shares for $20.2 billion in the quarter, and management intends to return roughly half of free cash flow to shareholders . That is a franchise using a de-rated multiple to shrink its own share count while it waits for the cadence question to answer itself.

The verdict

The market has taken the world’s most-owned stock and, inside two months, compressed it to roughly 22–24x consensus NTM / FY27 earnings — not clearly below the S&P 500, but strikingly close to a market multiple for a company whose revenue just grew 85% year over year — on a bear case built around a noisy prior-generation Hopper rental datapoint and an unconfirmed manufacturing rumor, even as NVIDIA says Blackwell is now the majority of revenue and its own July pricing slide showed H100 AI cloud GPU pricing up 13% over three months and 21% over six months. That is a sentiment rotation dressed as a fundamental re-rating. What the tape actually shows is a business whose second engine — the rack-scale network — is compounding faster than the compute it wraps, whose latest raise excludes China entirely, and whose margins held through a full architecture transition. The one genuine risk, a Rubin cadence slip, is dated to a print seven weeks out and is the reason the fair range is wide rather than a point. At roughly 23x forward earnings against a business still growing revenue north of 80% with the cash economics of a near-monopoly, the risk-reward has skewed to the upside, and the market is underpricing the networking leg that carries it. NVIDIA is worth more than the compute-marketplace narrative is willing to pay for it.

Disclaimer: Tenzing Analytics, Inc. is not a registered investment adviser, broker-dealer, or financial institution, and all content provided by Tenzing Memo is AI-generated for informational purposes only. It does not constitute investment advice, a recommendation, or an offer or solicitation to buy, sell, or hold any security or financial instrument. Content is general in nature, not tailored to any individual's financial situation or objectives, and may contain errors, omissions, or inaccuracies; it should not be relied upon as a primary basis for any investment decision. Any forward-looking statements are subject to risks and uncertainties and may not reflect actual future results. Users should conduct their own research and consult qualified financial, legal, or tax advisors before making investment decisions, and by accessing this content, you acknowledge and accept these limitations.