NVIDIA Customer Concentration: A Big 4 Earnings Preview

NVIDIA (NVDA) just closed fiscal 2026 with $216 billion in revenue, up 65% year over year, and $97 billion in free cash flow. By any measure, the business is operating at an unprecedented pace. NVDA’s disclosed backlog for Blackwell and Rubin doubled to $1 trillion through 2027, compared with the prior forecast through 2026. Yet daily announcements are coming in about customers’ plans to diversify away from NVDA.

Two things are apparent, and both are accelerating. NVIDIA’s customer base is becoming more concentrated. Alternatives that once lived on hyperscaler product roadmaps now operate at a gigawatt scale in production, and the largest AI labs are designing their own chips. Four sets of announcements over the past two weeks made both explicit:

• Cerebras’s S-1 filing on April 17, disclosing OpenAI’s $20 billion-plus inference compute commitment
• Amazon’s expanded Anthropic partnership on April 20
• Google Cloud Next 2026 (April 22-24), where Google previewed a dual-architecture eighth-generation TPU and reaffirmed roughly $180 billion in 2026 infrastructure capex
• Google’s $40 billion investment in Anthropic on April 24.

We built a model to put concrete numbers to each.

The underlying financial model

Tables 1 and 2 below were built with Daloopa Scout, covering:

• NVDA customer concentration figures and data center revenue by quarter
• Capex for MSFT, GOOGL, AMZN, META, and ORCL by quarter

With Scout, reconciling NVIDIA’s January fiscal year-end against the calendar-year hyperscalers took minutes rather than a day of manual 10-Q pulling, with every figure linked directly to its exact source in the filing. Table 3 draws directly on company announcements and SEC filings.

Concentration: from manageable to material

NVIDIA’s most recent disclosure shows four direct customers, each representing more than 10% of total revenue, collectively accounting for 61% of the business. The top single customer alone now represents 22%. A year ago, three customers each at roughly 12% added up to about 36%. The shape of the customer mix has materially changed over the past 12 months.

The accounts receivable mix is even more concentrated. The largest AR customer accounted for 25% of NVIDIA’s $38.5 billion AR balance at year-end, or roughly $9.6 billion. The second-largest accounted for another 18%.

Two things make this more than a “big company, big customers” story: First, NVIDIA’s data center segment now represents 92% of total revenue, up from roughly 60% two fiscal years ago. Concentration in DC equals concentration in the overall business. Second, those four customers are each running custom silicon programs at scale. The buyers behind the concentration are the same ones building alternatives — that’s the more consequential risk.

The capex pie is growing. NVIDIA’s slice is not.

The conventional bull case argues that the absolute size of the hyperscaler capex pool is what matters. Even if any individual hyperscaler builds custom silicon, the aggregate spend continues to flow through NVIDIA at scale. Hyperscaler capex hit $138.3 billion across MSFT, GOOGL, AMZN, META, and ORCL in Q4 2025, up 66% year over year. The pie is enormous and growing.

The capex-to-revenue spread tells a more nuanced story. We measured the difference between NVIDIA’s data center revenue growth and total hyperscaler capex growth across the past two years.

For two consecutive quarters in mid-2025, hyperscaler capex grew faster than NVIDIA’s data center revenue. Q4 2025 produced a recovery, but the underlying pattern is harder to dismiss: the share of incremental capex flowing to NVIDIA is no longer expanding. Over the past eight quarters, NVDA’s data center revenue has accounted for between 39% and 47% of total hyperscaler capex. The peak was 47% in Q1 2025. The latest is 45%.

The bull rebuttal is reasonable. Quarter-to-quarter timing of GPU shipments is lumpy. Networking revenue (up more than 3.5x year over year to nearly $11 billion in Q4) and sovereign AI demand (more than $30 billion for the year) are diversifying the customer mix. None of that is wrong. But spread compression is exactly what you would expect to see if a structural share shift were underway.

The custom silicon scoreboard, hyperscalers

From 2024 through early 2025, hyperscaler custom silicon was a roadmap story. The April 2026 news cycle has moved it firmly into the present tense. Going player by player:

Amazon. On April 20, Amazon committed up to $25 billion of additional investment in Anthropic, paired with Anthropic’s $100 billion, ten-year commitment to AWS Trainium and Graviton:

• 5 GW of Trainium capacity locked in for Anthropic alone
• Trainium plus Graviton combined run rate over $10 billion, growing at triple digits year over year
• Trainium2 fully subscribed at 1.4 million chips deployed; Trainium3 essentially sold out through mid-2026
• Project Rainier, the Anthropic training cluster, runs over 500,000 Trainium2 chips

Google. Four days later, Google announced up to $40 billion of investment in Anthropic along with 5 gigawatts of dedicated Google Cloud TPU capacity. That stacks on top of the 3.5 gigawatts of Broadcom-built TPU capacity Anthropic and Google announced earlier in April, scheduled to come online in 2027.

The bigger TPU disclosure came at Google Cloud Next 2026 in Las Vegas this week:

• Ironwood (TPU v7) generally available: 4.6 petaFLOPS per chip, 42.5 exaFLOPS in a 9,216-chip superpod; positioned as TPU’s inference-era flagship
• Eighth-generation TPU previewed — first to split training and inference into separate chip designs:
  • TPU 8t (Sunfish): Broadcom, training; 9,600 chips per superpod, 2 petabytes shared HBM, 2.7x performance per dollar versus Ironwood
  • TPU 8i (Zebrafish): MediaTek, inference; 3x on-chip SRAM of Ironwood, 288 GB HBM per chip
  • Both target TSMC 2nm and late 2027 deployment
• Capex reaffirmed at $175 to $185 billion for 2026, nearly double 2025’s $91 billion; just over half of Google’s ML compute will go to TPU over third-party GPUs; 4.3 million TPU shipments projected for 2026, 10 million in 2027, and more than 35 million in 2028

The split-chip strategy is the most consequential signal. Google is not just building one TPU. It is building two purpose-built chips with two different design partners (Broadcom for training, MediaTek for inference), and reportedly in talks with Marvell as a third. That is the architecture of a mature merchant-quality competitor to NVIDIA’s general-purpose GPU portfolio, not a captive science project. Anthropic is the disclosed anchor customer for both 8t and 8i.

Meta. Meta’s MTIA (Meta Training and Inference Accelerator) program disclosed two material updates in the past six weeks. On March 11, Meta laid out a four-generation roadmap — all targeted by the end of 2027 on roughly a six-month cadence:

• MTIA 300: Hundreds of thousands deployed in production; running internal recommendation models and tested with Llama
• MTIA 400: In lab validation, headed to data centers
• MTIA 450: Mass deployment scheduled for early 2027
• MTIA 500: Later in 2027

On April 14, Meta and Broadcom extended their partnership through 2029 to co-develop those generations. The initial commitment exceeds 1 gigawatt of MTIA capacity, with the potential to scale to multi-gigawatt deployments. Broadcom characterized the chips as the industry’s first AI accelerator built on a 2-nanometer process node. Hock Tan stepped down from Meta’s board on the same day to take an advisory role in Meta’s silicon strategy. Meta has separately committed to 6 gigawatts of AMD MI450 GPUs starting in late 2026, an estimated $60 billion deal that further compresses the addressable NVIDIA opportunity inside Meta.

Two features make Meta the most aggressive of the four hyperscalers when it comes to substitution. MTIA is purely for internal use (unlike TPU and Trainium, which are also rented externally), so every gigawatt of MTIA capacity is a direct displacement of what Meta would otherwise spend on NVIDIA. And Meta is shipping a new chip generation every six months, an unusually fast cadence for custom silicon.

Microsoft. Microsoft remains the most NVIDIA-aligned of the four. Public disclosure on Maia and Cobalt remains limited, OpenAI workloads inside Azure are still heavily NVIDIA-dominant, and the cadence of Microsoft’s in-house program looks slower than Amazon’s, Google’s, or Meta’s. The asterisk is that OpenAI itself is now building chips, so even Microsoft’s footprint has a structural path toward non-NVIDIA migration over time.

The chip race goes upstream: AI labs and a public-market merchant

The largest cloud computing customers are now bypassing cloud providers’ silicon and designing or buying their own at scale.

OpenAI. On October 13, 2025, OpenAI and Broadcom announced a strategic collaboration to deploy 10 gigawatts of OpenAI-designed AI accelerators. OpenAI handles chip design. Broadcom handles development, networking (Ethernet for both scale-up and scale-out), and deployment. TSMC handles manufacturing. Production starts in the second half of 2026 with full deployment targeted by the end of 2029. The chips are exclusively for OpenAI’s internal use. Combined with OpenAI’s separately announced AMD MI450 commitment (6 GW starting H2 2026, with AMD warrants potentially giving OpenAI roughly 10% of AMD’s equity) and the NVIDIA partnership ($100 billion of NVIDIA investment, 10 GW of NVIDIA-based systems), OpenAI’s 2027 compute footprint is triangulated across multiple suppliers.

That triangulation got a fourth leg on April 17, when Cerebras filed its S-1 for a Nasdaq IPO under the ticker CBRS. The filing disclosed a Master Relationship Agreement with OpenAI valued at more than $20 billion over multiple years, covering 750 megawatts of Cerebras wafer-scale inference compute through 2028, with options for an additional 1.25 gigawatts through 2030. OpenAI separately extended Cerebras a $1 billion working capital loan at 6% interest and received warrants for up to 33.4 million Cerebras shares, a structure that gives OpenAI a path to roughly 10% ownership if it consumes the full capacity. OpenAI and Cerebras have also agreed to co-design future models for future Cerebras hardware, a deep technical relationship that goes beyond a typical supply contract.

OpenAI’s disclosed 2027 to 2030 compute footprint now spans four silicon architectures, with three of the four not being NVIDIA:

• Broadcom-designed XPUs: 10 GW
• AMD MI450 GPUs: 6 GW
• Cerebras WSE inference systems: up to 2 GW
• NVIDIA-based systems: 10 GW

Anthropic. Reuters reported on April 9 that Anthropic is exploring the design of its own AI chips. The effort appears to be in its early stages. No engineering team has been formally assembled. No final architecture has been selected. However, Anthropic (which already runs Claude across Trainium, TPU, and NVIDIA, with explicit gigawatt commitments to all three), considering this path indicates where the cost curve has moved. Industry estimates put advanced AI chip development at $500 million or more, before software co-design, fabrication ramp-up, and ecosystem tooling. At a $30 billion run rate (up from roughly $9 billion at end of 2025), the math is starting to work.

Cerebras. The Cerebras filing matters beyond OpenAI’s specific commitment. It marks the moment a non-NVIDIA, non-hyperscaler merchant silicon vendor reached genuine commercial scale and chose to test the public market. Cerebras reported $510 million in 2025 revenue, up 76% year over year, and swung to net income of $87.9 million from a $485 million loss the prior year. The company’s WSE-3 wafer-scale chip integrates 4 trillion transistors and 900,000 AI cores onto a single wafer, an architecture that Cerebras positions as structurally advantaged for inference latency. In March 2026, Cerebras announced a partnership with AWS to deploy the first non-AWS-built silicon at scale in AWS data centers, pairing Trainium for prefill processing with Cerebras WSE for decode in a disaggregated inference architecture. Cerebras still carries a meaningful concentration of its own (G42 and MBZUAI accounted for 86% of 2025 revenue), but the OpenAI and AWS contracts fundamentally reshape the customer mix going forward. AMD’s parallel deals with Meta (6 GW MI450) and OpenAI put it on a similar trajectory. The takeaway: NVIDIA’s competitive set has broadened from captive in-house programs to public-market merchant alternatives, with distribution across both AI labs and hyperscalers.

The strategic implication for NVIDIA is structural. Through 2025, the substitution thesis ran through the four hyperscalers. If you believed Microsoft, Amazon, Google, and Meta would each build chips at scale, you were bearish. If you didn’t, you weren’t. The 2026 version is broader. The substitution thesis now runs through every customer at a sufficient scale, including the AI labs themselves, and is reinforced by merchant alternatives like Cerebras and AMD that are reaching commercial relevance. By 2027, the only large-scale AI workloads guaranteed to run on NVIDIA will be those inside enterprises and sovereigns that lack the scale or talent to build, fund, or migrate to an alternative.

What survives the squeeze

The bull case still has real components.

Networking and systems revenue is the most underappreciated piece of the story. Even when a hyperscaler swaps GPUs for in-house accelerators, much of the rack infrastructure (Spectrum-X Ethernet, NVLink, InfiniBand) remains NVIDIA. Q4 networking revenue of $11 billion grew more than 3.5x year over year for a reason. CUDA’s incumbency advantage on the ISV and developer side is real and not closing quickly. Sovereign AI demand is a structurally new pool, less correlated with US hyperscaler spend, and growing fast enough to absorb a meaningful share of supply through 2027. Q1 FY27 guidance of $78 billion (up 15% sequentially) signals that, in the near term, demand is still outrunning supply.

Even Google’s dual-chip TPU v8 announcement uses Broadcom Ethernet, not NVLink, and Meta’s MTIA racks use Broadcom networking too. The scale-up and scale-out fabric layer is itself a contested market now, not a NVIDIA stronghold. That is a quieter erosion than the headline GPU-substitution story, but it matters for the networking segment’s longer-term margin profile.

That demand is outpacing supply is the binding constraint right now. The interesting question is what happens when it is no longer binding.

The unresolved question

Two years ago, the NVIDIA bull case was simple: AI infrastructure is a generational platform shift, and NVIDIA owns it. That call was correct.

The 2026 bull case must be more specific. AI infrastructure remains a generational platform shift, but NVIDIA now must share it with three hyperscalers running multi-gigawatt custom silicon programs (Amazon, Google, Meta), one hyperscaler (Microsoft) whose largest AI tenant is itself building chips, two AI labs (OpenAI committed to 10 GW Broadcom plus up to 2 GW Cerebras, Anthropic exploring), at least one merchant alternative now scaling toward a public listing (Cerebras). It continues to derive 61% of its revenue from a customer base in which each major buyer also funds the alternative.

Disclosed non-NVIDIA commitments through 2030:

• Anthropic: 10 GW
• OpenAI: 18 GW (10 GW Broadcom, 6 GW AMD, up to 2 GW Cerebras)
• Meta: multi-GW MTIA plus 6 GW AMD MI450
• AWS Bedrock customers: multi-GW Trainium
• Google (Gemini and external): multi-GW TPU; 35 million shipments projected by 2028

It does not have to fall to zero, or even to half of its current level, for the equity story to change materially. The current valuation embeds an expectation that NVIDIA continues to capture roughly the same 45% share of a much larger hyperscaler capex pie. The April 2026 announcements suggest the pie is still growing, but the slice is starting to narrow at the edges, and the base of credible alternative suppliers is now broader than at any point since the AI build-out began.

What to watch when the Big 4 report earnings this week

Microsoft, Alphabet, Amazon, and Meta all report on April 29. The investor-relevant questions are not about beat or miss. They are about silicon mix.

• Alphabet: whether the $175 to $185 billion 2026 capex guide is revised; Ironwood ramp disclosure through 2026; production timing for TPU 8t and 8i
• Amazon: updated Trainium plus Graviton run rate (last disclosed at over $10 billion); OpenAI utilization of Trainium beyond the 2 GW already disclosed; Trainium4 timing
• Meta: MTIA 400 deployment timing; AMD MI450 first-gigawatt readiness; in-house silicon as a share of 2026 inference compute
• Microsoft: Maia volumes; OpenAI utilization of Broadcom-designed chips in Azure; capex split between merchant silicon and in-house programs

NVIDIA’s next print is in late May, but the hyperscaler commentary on April 29 will tell you most of what you need to know about the trajectory of NVIDIA’s customer concentration and the pace at which custom silicon is moving from headline to deployed revenue. Outside the Big 4 prints, watch the Cerebras (CBRS) IPO pricing window. The pricing will be a real-time signal of public-market appetite for non-NVIDIA AI silicon, and the demand book will show whether institutional investors are willing to pay up for the same substitution thesis that NVIDIA bulls are still discounting.