Can it hold margins as its largest customers ship their own chips?

Thesis

NVIDIA is posting its best margins while its largest customers build the products that could take them away. In the quarter ended July 26, 2026 (the second quarter of fiscal 2027), revenue was $96.2 billion, up 106% from a year earlier, and GAAP gross margin was 75.0%. Hyperscale customers, the group that includes every company designing its own accelerator, bought $49 billion of the $89.0 billion Data Center total, about 55% (our arithmetic).

The margin pressure visible today does not come from custom silicon. Management expects gross margin to fall to 74% in the third quarter, bottom at 71% to 72% in the fourth, and settle at 72% to 73% in fiscal 2028, and it attributed the drop to memory prices rising faster than planned. At the third-quarter revenue guide of $108.0 billion, each point of gross margin is worth about $1.1 billion of quarterly gross profit, so a slide from 75% to 71% costs roughly $4.3 billion a quarter (our arithmetic).

Custom silicon is a real threat, and competitors’ numbers show it. Broadcom’s AI semiconductor revenue reached $16.7 billion in its latest quarter, up 221%, and Amazon says its in-house chip unit has an annual revenue run rate above $20 billion. Those chips compete for the same budgets. They have not yet forced a price concession that shows up in NVIDIA’s reported margin.

NVIDIA is responding by selling more of the data center per dollar of customer spend. Management says its revenue opportunity per gigawatt of capacity has grown from about $18 billion with Hopper to $25 billion with Blackwell and $40 billion with Vera Rubin. That content growth, plus a second customer group that does not design chips, is how the company intends to keep dollars per rack high even if the GPU itself faces price competition.

Margins look defensible in the near term and contestable beyond fiscal 2028. The evidence that would change that view is laid out under What Would Prove Us Wrong.

Company History

NVIDIA was incorporated in California in April 1993 and reincorporated in Delaware in April 1998. It began as a graphics chip company serving PC gaming, and the parallel processing it built for rendering became the base for its data center business. The company later introduced CUDA, a software platform that let developers run general-purpose workloads on its GPUs. That software layer is the reason NVIDIA’s installed base became sticky, because AI frameworks were written against it first.

The scale change is recent. Revenue was $60.9 billion in fiscal 2024, $130.5 billion in fiscal 2025 and $215.9 billion in fiscal 2026 (ended January 25, 2026). Headcount reached about 42,000 employees in 38 countries at that fiscal year end, of whom 31,000 work in research and development.

The company’s China position has shrunk to almost nothing. In the first quarter of fiscal 2026 it took a $4.5 billion charge on H20 excess inventory and purchase obligations after export restrictions tightened, and it describes itself in the 10-K as effectively foreclosed from China’s data center compute market. In the latest quarter, less than 1% of Data Center revenue came from Hopper 200 products shipped to customers in China, and the forward outlook assumes none.

The more recent strategic move is the shift from selling chips to selling systems. The company has added CPUs, three kinds of networking and, through a partnership announced last year, Groq’s inference processor technology to the rack. Management now describes the product as a full-stack AI factory platform rather than an accelerator.

Product

Data Center accelerators and systems

Blackwell is the current volume platform and Vera Rubin is the next one. NVIDIA began production shipments of Vera Rubin in August 2026, says it has purchase orders from every major hyperscaler, AI cloud and system maker, and expects Rubin to be about 20% of Data Center revenue in the third quarter. The company claims Vera Rubin delivers 30 times the throughput per megawatt and 35 times lower token cost than Grace Blackwell Ultra. These are vendor claims and we have not independently tested them.

Data Center revenue was $193.7 billion in fiscal 2026. Starting in the second quarter of fiscal 2027, the company reports Data Center in two parts: hyperscale, and ACIE (neoclouds, industrial and enterprise customers, including sovereign programs).

Networking

Networking is the quiet second engine. Management called the second quarter a record, with revenue up 18% sequentially, and said Spectrum-X Ethernet grew 2.6 times year over year. The company did not disclose a dollar figure for networking in the release we reviewed. The 10-K names Arista, Broadcom, Cisco, Marvell and others as networking competitors. Networking matters to the margin question because it ties customers to NVIDIA’s rack design even when the accelerator inside could be swapped.

CPUs

Grace, NVIDIA’s Arm-based CPU, generated more than $5 billion of trailing twelve-month revenue. The successor, Vera, is in full production and is also sold stand-alone. Management sees demand for about $20 billion of total server CPUs and expects CPU revenue to more than double in fiscal 2028. The 10-K flags Amazon, Huawei and Microsoft as Arm-based CPU competitors.

Inference with Groq

NVIDIA introduced the Groq 3 LPX rack-scale inference system in August 2026 and said it expects volume shipments later in the quarter, with Nebius first. This is the direct answer to inference-specific custom chips. It lets a customer bolt a low-latency accelerator onto a Vera Rubin deployment instead of buying a separate vendor’s part.

Gaming, visualization, automotive and edge

The smaller segments are still large in absolute terms. Fiscal 2026 revenue was $16.0 billion in Gaming, $3.2 billion in Professional Visualization and $2.3 billion in Automotive. In the latest quarter the company reports a combined Edge Computing line of $7.2 billion, up 27% year over year.

Market

Customer

Customers fall into two groups with different relationships to the margin question.

The first is the hyperscalers: Amazon, Microsoft, Alphabet, Meta, Oracle and a few peers. They bought $49 billion of Data Center product in the quarter, growing 13% sequentially. They are NVIDIA’s largest buyers and also its most capable would-be substitutes. The 10-K shows the concentration at the direct-customer level: in fiscal 2026 one direct customer was 22% of total revenue and another was 14%.

The second group is ACIE. Its revenue was $40 billion, up 25% sequentially and 138% year over year. Management says neocloud partners are expected to exit the year with 8 gigawatts of installed capacity, up from about 3 gigawatts at the end of 2025. Sovereign business, mostly through regional neoclouds, grew 35% sequentially. Management argues these buyers do not design chips and need a whole factory platform, which is the segment where NVIDIA’s pricing is least exposed to an in-house alternative. The third quarter growth guide leans on this group, and hyperscale growth is expected to reaccelerate in the fourth quarter and into fiscal 2028 as Rubin supply grows.

A third, cross-cutting group is the frontier AI labs. NVIDIA expects demand from the labs for which it will draw on its balance sheet to contribute roughly a quarter of its business next year. OpenAI’s existing and planned commitments represent about 12 gigawatts of NVIDIA compute.

Market Size

NVIDIA’s market is the capital spending of the companies that build AI capacity. On the August call, management said cloud industry backlog is greater than $2 trillion and that capital expenditure by the top five hyperscalers is expected to reach nearly $800 billion in 2026 and $1.3 trillion in 2027. These are management’s citations of industry expectations, not NVIDIA forecasts.

Venture funding adds a second demand source. Management cited global AI venture funding of more than $400 billion in the first half of 2026, about 70% of which is spent on compute, versus $265 billion raised in all of 2025.

NVIDIA’s own outlook is supply-limited. It expects fiscal 2028 revenue to grow approximately 70%, although customer forecasts imply growth doubling, and it expects supply to remain a bottleneck at least through the end of fiscal 2028. A supply-limited vendor holds pricing power. The question is what happens when supply catches up and buyers have alternatives.

Competition

Competitive Landscape

Competition comes from three directions. Merchant GPU rivals, chiefly AMD, sell a similar product to the same customers. Customer-designed accelerators, built with partners such as Broadcom, replace NVIDIA parts inside a single company’s data centers. Networking and CPU specialists attack individual layers of the NVIDIA rack. The 10-K lists the cloud providers Alibaba, Alphabet, Amazon, Baidu, Huawei and Microsoft among competitors with internal designs, and states that some customers have in-house chip development that could replace NVIDIA products.

The evidence so far shows coexistence. Anthropic is the clearest case. In October 2025 it said it would expand to up to one million Google TPUs, bringing well over a gigawatt online in 2026, and in the same post said it would keep using a mix of TPUs, Amazon Trainium and NVIDIA GPUs. A buyer with three suppliers can negotiate, and that is the mechanism by which custom silicon could reach NVIDIA’s pricing before it reaches NVIDIA’s volume.

Management’s response on the August call was that many XPUs are inference-specific chips for one cloud or one service, whereas NVIDIA sells a platform that runs in every cloud. That is a fair description of today’s product. It is also a claim that holds only while workloads remain varied and portable.

For scale: Broadcom’s AI semiconductor revenue of $16.7 billion is equal to about 19% of NVIDIA’s $89.0 billion Data Center revenue (our arithmetic, and Broadcom’s figure includes networking). Broadcom guides next quarter’s AI revenue to $21.7 billion, up 236%, so the gap is closing in relative terms from a low base.

Competitors

Broadcom and its custom accelerator customers

Broadcom designs custom AI accelerators (“XPUs”) for hyperscalers and labs. Its fiscal third quarter (ended August 2, 2026) brought revenue of $29.6 billion, up 86%, of which AI semiconductors were $16.7 billion. OpenAI and Broadcom unveiled the Jalapeño inference processor in June 2026, with initial deployment slated for late 2026. OpenAI says final performance testing is still underway, so its claims of better performance per watt are unverified. OpenAI is also a very large NVIDIA customer, which shows how entangled these relationships are.

Amazon Trainium

Amazon is the largest in-house program by disclosed revenue. On its first-quarter 2026 call, CEO Andy Jassy said the custom chip unit, covering Trainium, Graviton and Nitro, has a run rate above $20 billion and more than $225 billion in Trainium revenue commitments. He called it “one of the top three data center chip businesses in the world” and said Trainium2 offers about 30% better price-performance than comparable GPUs, an Amazon claim. Yet on the NVIDIA call, management said AWS will deploy an additional 2 million NVIDIA GPUs through the second quarter of fiscal 2029. The largest in-house chip builder is still expanding its NVIDIA purchases.

Google TPU

Google’s TPUs are the most mature merchant-adjacent alternative, now used at scale by Anthropic. We could not verify from a primary source reports that Google sells TPU hardware directly, and we do not rely on them.

AMD

AMD’s second quarter 2026 revenue was $11.5 billion, with its Data Center segment at $6.7 billion, up 107%. AMD announced a partnership to deploy up to 2 gigawatts of MI450 Series GPUs with Anthropic. AMD’s gross margin was 54% GAAP for the whole company, which gives a sense of the margin gap a merchant rival works at, though segment margins are not disclosed.

Huawei and China-based rivals

The 10-K names Huawei in accelerators, CPUs and networking. With China data center revenue at effectively zero in NVIDIA’s outlook, this competition affects the market NVIDIA has lost rather than the one it is defending.

Business Model

NVIDIA designs chips and does not fabricate them, so revenue converts to gross profit at software-like rates for a hardware company. It sells through direct customers, which include system makers and cloud providers, and increasingly sells reference designs for entire facilities. Management frames revenue per gigawatt as the unit of its model, with the figure rising from about $18 billion with Hopper to $40 billion with Vera Rubin.

Pricing power rests on three things: the CUDA software base, system-level co-design across GPU, CPU and networking, and priority access to scarce components such as high-bandwidth memory. The first two are structural. The third is the one currently squeezing margin. Management said memory price increases exceeded prior expectations and are heading higher into next year, and that it plans to recover part of the cost through executed price increases that take effect in the first quarter of fiscal 2028.

Capital intensity is low for the operating business: capital expenditure was $2.7 billion against revenue of $96.2 billion in the latest quarter. What has grown is the use of the balance sheet as a sales tool. NVIDIA announced financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR intended to raise over $500 billion of third-party capital for lab infrastructure, and said it will provide credit enhancement on nearly 2 gigawatts of compute for another frontier lab. Management acknowledged that critics will call this circular financing. We treat it as a real second-order risk, covered below.

Financials

Latest quarter

Revenue in the second quarter of fiscal 2027 was $96.2 billion, up 18% from the prior quarter. Data Center contributed $89.0 billion, or about 92.5% of the total (our arithmetic). Gross margin was 75.0% on both a GAAP and non-GAAP basis, compared with 72.4% GAAP a year earlier.

GAAP operating income was $63.7 billion on GAAP operating expenses of $8.4 billion. GAAP net income was $59.7 billion, or $2.46 per diluted share. That figure includes $7.8 billion of other income, almost all gains on equity securities. Non-GAAP net income, which excludes those gains, was $54.0 billion, and the cleaner comparison to operating results is the non-GAAP figure.

Cash flow and balance sheet

Operating cash flow was $24.1 billion and free cash flow $21.3 billion, well below net income and below the prior quarter’s $48.6 billion. The weak conversion reflects working capital. Inventory rose to $31.6 billion ahead of the Vera Rubin launch, receivables were $63.1 billion, and days sales outstanding rose to 60 days because large investment-grade customers received extended payment terms on purchases that ship over several quarters. Over six months, free cash flow was $69.9 billion, which is still strong.

Cash and marketable securities were about $99.4 billion (our sum of $22.4 billion of cash, $34.1 billion of debt securities and $42.8 billion of equity securities). Non-marketable securities, largely stakes in the companies NVIDIA also sells to, stood at $51.2 billion. Long-term debt rose to $32.4 billion from $7.5 billion at the January 2026 year end.

The company returned a record $26.0 billion to shareholders in the quarter, with about $99.0 billion remaining on its repurchase authorization. On September 28 the board added $150 billion, raising the total remaining authorization to $235 billion.

Guidance

Third-quarter revenue is guided to $108.0 billion, plus or minus 2%, assuming no China Data Center compute revenue. Gross margin is guided to 74.0%, plus or minus 50 basis points. Non-GAAP operating expenses are guided to about $9.0 billion, and management expects full-year operating expense growth in the low 50s percent. The tax rate is expected to be 16% to 18%.

Fiscal 2026 for context

Fiscal 2026 revenue was $215.9 billion, GAAP gross margin 71.1% (against 75.0% in fiscal 2025, depressed by the H20 charge), GAAP net income $120.1 billion and free cash flow $96.6 billion. Trailing twelve-month revenue to July 26, 2026 is about $303.0 billion (our computation: the third and fourth quarters of fiscal 2026, $57.0 billion and $68.1 billion, plus the first and second quarters of fiscal 2027, $81.6 billion and $96.2 billion).

Valuation

As of the October 30, 2026 close, NVIDIA’s market capitalization was about $4.93 trillion (our computation). That is the closing price of $204.64, down 4.2% from $213.66 the day before, times 24.1 billion shares outstanding on the cover of the 10-Q for the quarter ended July 26, 2026 (count as of August 21, 2026). The highest close in the 2026 daily price history we reviewed was $239.24 on October 6, which on the same share count is about $5.77 trillion, so the equity is roughly 14% below that peak (our arithmetic).

On trailing twelve-month revenue of about $303.0 billion, the market value equals roughly 16.3 times sales (our computation). On fiscal 2026 diluted GAAP earnings of $4.90 per share, the share price equals roughly 41.8 times. That ratio overstates the current multiple. Diluted GAAP earnings for the last four quarters were $7.91 per share (our sum of $1.30 and $1.76 for the third and fourth quarters of fiscal 2026 and $2.39 and $2.46 for the first and second quarters of fiscal 2027), which puts the price at roughly 25.9 times trailing earnings. Those earnings include large gains on equity securities.

Against its own history, the multiple has compressed. At the end of 2024 the market value was $3.29 trillion against fiscal 2025 revenue of $130.5 billion, about 25 times sales (our approximation, as fiscal and calendar years differ by a few weeks). At the end of 2025 it was $4.53 trillion. The market value rose 38% over those two year-ends while revenue on a fiscal basis rose 65% in fiscal 2026 alone, so growth has outpaced the equity.

Balance sheet context matters: about $99.4 billion of cash and securities against about $33.4 billion of debt (the $32.4 billion long-term plus $1.0 billion short-term), though a large share of those securities are equity stakes whose value moves with markets. The market is capitalizing a margin structure that management itself guides lower for two quarters. If gross margin settles at 72% to 73% rather than 75%, the same revenue yields lower profit, and the multiple on forward earnings depends on that outcome as well as on growth.

We make no claim about fair value. This section describes where the market prices the company and what it is pricing in.

Key Opportunities

Non-hyperscale demand that does not design chips

ACIE already generates $40 billion a quarter and grew 138% year over year. Management expects it to drive third-quarter sequential growth and to account for roughly half of Data Center over time. This is the strongest answer to the margin question, because a sovereign cloud or enterprise buyer comparing platforms has no custom chip to threaten the price. The risk is that some of this growth is neoclouds financed by NVIDIA-linked capital (see Key Risks), which makes it less independent than it looks.

More content per gigawatt

Raising revenue per gigawatt from $25 billion to $40 billion by adding CPUs, networking and Groq inference racks means the company can concede price on an individual accelerator and still grow dollars per site. A custom ASIC replaces one component. The Vera, Spectrum-X and NVLink content around it has fewer substitutes. The CPU line alone, with trailing revenue above $5 billion and expectations of more than doubling in fiscal 2028, shows the approach working in a new category.

Balance-sheet-backed lab and infrastructure financing

The financing platforms aim to raise over $500 billion for AI factories, and the SoftBank Energy Portsmouth site is planned to support 4.25 gigawatts for OpenAI on NVIDIA compute. Locking in labs through multi-year site commitments reduces the chance that those labs switch architectures mid-build, which is where custom silicon would otherwise win share.

Key Risks

Custom silicon starts to set prices

The risk to margin is negotiated price. Amazon’s $225 billion of Trainium commitments, Broadcom’s guided $21.7 billion AI quarter, Anthropic’s up to one million TPUs and OpenAI’s own inference processor give buyers a credible alternative at the table. Hyperscalers represent about 55% of Data Center revenue ($49 billion of $89.0 billion, our arithmetic). A customer that can move even a fraction of inference to its own part gains bargaining power over the rest. Because supply is the binding constraint today, that power is dormant. It becomes active when supply catches up, which management does not expect before the end of fiscal 2028.

Input costs and supply

Memory is the visible margin risk. Management guided gross margin down to 71% to 72% in the fourth quarter and said memory costs were heading higher. The recovery to 72% to 73% depends on price increases being accepted by customers who are simultaneously building alternatives. Inventory of $31.6 billion and large supply commitments mean a demand miss would hit the income statement quickly.

Concentration and customer credit

Two direct customers were 36% of fiscal 2026 revenue (22% plus 14%). Receivable days have risen to 60 as payment terms stretch. NVIDIA also holds $51.2 billion of non-marketable investments and offers credit support for labs. If a major lab cannot raise capital, NVIDIA could face revenue, receivable and investment losses at once. Management calls the compute redeployable, which limits but does not remove the exposure.

Customer capital spending

The demand case depends on top-five hyperscaler capital expenditure rising from nearly $800 billion in 2026 to $1.3 trillion in 2027. These are management’s figures. Any pause in spending reduces both volume and the scarcity that supports price.

China and export policy

NVIDIA’s outlook assumes no China data center compute revenue, and the 10-K says the company is effectively foreclosed from that market. An H200 license was granted in February 2026 but had produced no revenue at the filing date, and H200s face a 25% tariff on importation to the U.S. Policy could reopen the market, and management gives no timing.

What Would Prove Us Wrong

Our thesis is that NVIDIA can defend a gross margin in the low 70s through fiscal 2028 despite growing custom silicon. These observations would falsify it.

1. Gross margin below the guided floor. Any quarter of fiscal 2028 (February 2027 to January 2028) with GAAP gross margin below 71%, the bottom of the 71% to 72% range guided for the fourth quarter of fiscal 2027, would show price increases are not holding.
2. Hyperscale growth fails to reaccelerate. Management expects hyperscale growth to pick up in the fourth quarter. If hyperscale revenue grows less than the 13% sequential rate of the second quarter in the quarter ending January 2027, with supply no longer the explanation, in-house chips are taking share.
3. Custom accelerator revenue closes the gap. Broadcom’s AI semiconductor revenue was about 19% of NVIDIA’s Data Center revenue in the latest comparison. If Broadcom’s AI revenue exceeds 25% of NVIDIA’s Data Center revenue for the quarters ending around January 2027, custom silicon is scaling faster than the thesis allows.
4. Receivables and credit support grow faster than revenue. If days sales outstanding exceed 75 days (from 60) at the fiscal 2027 year end in January 2027, the sales are being financed more than they are being earned.

Summary

NVIDIA reported $96.2 billion of revenue and a 75.0% gross margin in its latest quarter, and guides $108.0 billion for the next. Market capitalization stood near $4.93 trillion at the October 30, 2026 close, about 16 times trailing sales (our computation).

Near term, memory costs drive the guided decline to 71% to 72%, and supply constraints keep customers from negotiating hard. Longer term, hyperscalers that account for about 55% of Data Center revenue are scaling Trainium, TPU and Broadcom-built accelerators, and OpenAI, Anthropic and Amazon each use more than one supplier. NVIDIA’s defense is a wider rack (CPUs, networking, inference racks) and a second customer group, ACIE, that buys full platforms.

We would revise toward lower margins if gross margin breaks below the guided floor, hyperscale growth lags once supply eases, or custom accelerator revenue rises above a quarter of NVIDIA’s Data Center sales. We would revise toward more durable margins if the first-quarter fiscal 2028 price increases hold and ACIE keeps growing faster than hyperscale.