← Back to blog
Essay No. 020  ·  AI Infrastructure  ·  Melbourne, Australia
AI Infrastructure Nvidia AI Capex Semiconductors Data Centers GPUs Neoclouds OpenAI CoreWeave TSMC ASML HBM Earnings Quality

Nvidia’s Earnings Quality Test.Original analysisNot investment advice

The numbers are real. The question is who ultimately pays for the AI buildout.
PM
Pugalenthi Magendran
March 2026  ·  Melbourne, Australia
12 min read

In 2021, Nvidia’s earnings-quality question was crypto. In 2026, it is AI capex. Nvidia’s revenue is real, but the harder question is whether today’s AI infrastructure demand turns into durable utilization, customer cash flow, and profitable compute consumption.

In 2021, the Nvidia earnings-quality question was crypto. Nvidia was growing fast. Gaming was booming. Data Center was becoming the future. But there was a concern underneath the headline numbers. How much of the growth was coming from real gaming demand, and how much was being inflated by crypto mining?

The uploaded 2021 SemiAnalysis piece captured that moment well. Nvidia had just reported strong results, and the article argued that CMP mining GPU sales were material and potentially temporary, with the risk that crypto demand could be inflating current earnings without representing durable long-term growth.1

That was the old earnings-quality problem. In 2026, the numbers are much bigger. Nvidia is no longer dealing with a few hundred million dollars of CMP revenue. It is reporting tens of billions in quarterly Data Center revenue and hundreds of billions in annual revenue. This is not a crypto side pocket. This is the centre of the AI economy.

But the earnings-quality question has not disappeared. It has changed.

The question is no longer how much of Nvidia’s gaming revenue is secretly crypto. The new question is how much of Nvidia’s AI revenue is durable end-user demand, and how much depends on concentrated customers, financed AI infrastructure, neocloud balance sheets, circular partnerships, and hyperscaler capex staying high.

Key idea

The correct claim is not that Nvidia’s earnings are fake. The correct claim is that Nvidia’s earnings are real, but the quality of those earnings depends on whether AI infrastructure capex turns into durable utilization, customer cash flow, and profitable AI usage.


I. The 2021 thesis

In May 2021, Dylan Patel published a SemiAnalysis piece on Nvidia’s reported quarter. The framing was not that Nvidia was a bad company. It was an earnings-quality question. CMP mining GPUs were booked into OEM & Other. Crypto miners may have inflated gaming GPU demand as well. Ethereum was migrating toward proof-of-stake. A used-GPU flood was a plausible risk if mining demand collapsed.1

I revisited that piece because the analytical move it made aged well. Separate durable demand from temporary cycle distortion. Ask what happens when the distortion fades. The 2026 distortion is different in shape, but the analytical question is the same.

2021 thesis

Temporary demand can look like structural growth until the cycle turns. In 2021, the distortion was crypto mining. The earnings-quality question was how much demand would remain after the crypto cycle faded.


II. The 2026 update

The distortion question is no longer crypto. It is AI infrastructure finance and utilization. The risk is not that miners disappear. The risk is that some of the AI capex behind today’s revenue is being pulled forward by customers racing to build capacity before the unit economics are fully proven.

2021 framing

Crypto-distorted demand

  • Risk · crypto inflated gaming and CMP demand.
  • Trigger · Ethereum proof-of-stake, crypto-price crash.
  • Failure mode · used-GPU flood, weaker new-GPU demand.
  • Question · will miners disappear?
2026 framing

AI capex pulled forward

  • Risk · AI capacity built ahead of monetised usage.
  • Trigger · hyperscaler restraint, neocloud stress, export controls.
  • Failure mode · underused GPUs, weak repeat orders, margin pressure.
  • Question · will deployed compute be utilised and paid for?

III. The earnings are real

The first move in a serious earnings-quality analysis is to take the numbers seriously. Nvidia reported Q4 FY2026 revenue of $68.1B, with Data Center revenue of $62.3B in the quarter, and full-year FY2026 revenue of $215.9B.2 That is not a temporary side product. That is one of the largest infrastructure buildouts in technology history.

Chart · Nvidia FY2026 revenue scale
Full year FY26 revenue
$215.9B
Q4 FY26 revenue
$68.1B
Q4 FY26 Data Center
$62.3B
Q4 Data Center revenue alone equals roughly 29% of full-year revenue. Source: Nvidia Q4 and FY2026 financial results.2

So the mistake would be to call Nvidia’s earnings fake. The better question is quality. A high-quality revenue dollar comes from a customer with durable demand, strong cash flow, high utilisation, and a repeatable economic reason to buy more. A lower-quality revenue dollar may still be real revenue, but depends on weaker assumptions: future financing, speculative capacity, low utilisation, customer concentration, or supplier-funded ecosystem loops.

The mistake would be to call Nvidia’s earnings fake. The better question is quality.


IV. The new issue is customer concentration

The first quality concern is concentration. Nvidia’s own Q3 FY2026 10-Q discloses that four direct customers each represented more than 10% of total revenue, with shares of roughly 22%, 15%, 13%, and 11%. The filing notes that revenue continues to be concentrated among a limited number of direct, indirect, and cloud-service purchasers, primarily attributable to Compute & Networking.3

Diagram · Top four direct customers, Q3 FY2026
22%
Customer A
15%
Customer B
13%
Customer C
11%
Customer D
~61% of total revenue attributable to four direct customers in the quarter, with the remainder spread across the broader buyer base.3
The AI boom is broad in narrative. Nvidia’s revenue path is still narrow in customer concentration.

This is not automatically bad. AI infrastructure is expensive, so the buyers are naturally huge. But if a few large customers slow orders, delay data-center builds, shift to internal ASICs, or hit power constraints, Nvidia feels it quickly.

The AI boom is broad in narrative. Nvidia’s revenue path is still narrow in customer concentration.


V. The circularity question

The second quality concern is the shape of the ecosystem deals. Nvidia and OpenAI announced a strategic partnership for at least 10 GW of AI data centers using Nvidia systems, with Nvidia stating its intent to invest up to $100B in OpenAI progressively as each gigawatt is deployed, with the first gigawatt planned for the second half of 2026 on Vera Rubin.45

Reuters has also reported on the depth of the Nvidia / CoreWeave relationship, including a roughly $2B Nvidia investment in CoreWeave and a multi-year commitment under which Nvidia agreed to purchase unsold cloud capacity from CoreWeave for up to $6.3B through 2032, with investor commentary raising questions about circular financing in the AI ecosystem.6

Diagram · The circularity question
Step 01
Nvidia sells GPUs, networking, and full systems to customers (hyperscalers, neoclouds, AI labs).
Step 02
Nvidia invests capital into selected customers (OpenAI, CoreWeave, others).
Step 03
Customer buys more Nvidia systems and / or offers Nvidia-backed compute capacity.
Question
Independent end-user demand, or supplier-supported ecosystem growth? Not fraud. An earnings-quality question.
A supplier investing in its customers can accelerate adoption. It can also blur the line between independent demand and ecosystem-funded demand.46

VI. The margin question

The third quality concern is what happens to margins as Nvidia sells more systems rather than more chips. Nvidia’s own commentary on the FY2026 transition from Hopper HGX systems to Blackwell full-scale data-center solutions, alongside an H20 inventory and purchase-obligation charge, sat in the middle of an otherwise strong margin profile.2 The structural question is what the cost stack looks like when the unit being sold is closer to an AI factory than a single accelerator.

Diagram · Cost stack inside the AI factory
10
Software, NIM, NeMo, AI Enterprise
High margin
09
GPU silicon
High margin
08
CPU + DPU + NVLink
Margin moat
07
HBM & high-bandwidth memory
Cost pressure
06
CoWoS / advanced packaging
Cost pressure
05
Substrates, boards, switches, optics
System cost
04
Racks, power, cooling, NVL system build
Logistics
03
Test, integration, customer config
Execution
02
Inventory, tariffs, purchase obligations
Risk
01
Foundry capacity at TSMC
Floor
The bigger the system Nvidia sells, the more revenue it can capture. The harder it becomes to keep chip-like margins on system-like products.

The bigger the system Nvidia sells, the more revenue it captures. The harder it becomes to keep chip-like margins on system-like products.


VII. The China question

The fourth quality concern is geography. US export controls have required licenses for H20 exports to China, and Nvidia’s commentary has discussed a meaningful inventory and purchase-obligation charge tied to that product line, with the company also framing forward outlook assumptions that exclude Data Center compute revenue from China.2

That is the unusual part of Nvidia’s 2026 setup. The company can be demand-constrained and geopolitically constrained at the same time. Demand can be massive and still not fully addressable.

Demand can be massive, and still not fully addressable.


VIII. The physical bottleneck question

The fifth quality concern is the physical world. Nvidia’s revenue depends on the rest of the AI factory stack actually showing up: TSMC advanced wafers, HBM supply, CoWoS and advanced packaging, substrates, networking silicon, optics, power delivery, liquid cooling, transformers, land, grid connections, and data-center construction. TSMC’s own 2025 annual report frames AI/HPC as a structural driver of advanced process demand and packaging investment, while ASML’s 2025 strategic report makes a parallel point: AI requires leading-edge processors and a significant increase in DRAM compared with traditional compute architectures.78

The AI factory is not just a product. It is a capital project.


IX. Utilization is the final test

The final quality concern is what happens after the systems are deployed. Buying GPUs is not the end of the story. The installed compute must be used, and the usage must produce enough revenue to justify the next order.

Diagram · The utilization funnel
GPU purchase
Customer signs an order with Nvidia.
Data-center deployment
Site built, power and cooling secured, systems racked.
Utilisation
Systems run real training or inference workloads at scale.
AI product revenue
Outputs feed paying customers, agents, enterprise tools, or research.
Customer cash flow
Revenue covers depreciation, power, hosting, and software costs.
Repeat orders
Customer returns with internally funded demand, not supplier financing.
Capex is the purchase. Utilisation is the proof. Earnings quality lives in the gap between the two.

X. What could break the bear case

This section has to be honest. The bull case for Nvidia in 2026 is not based on hype. It rests on a set of arguments that are individually defensible.

Bull case

Why the cycle may keep going

  • AI usage is still growing. Token volumes, enterprise pilots, and agent deployments expand.
  • Full-stack platform. CUDA, networking, and rack-scale systems are hard to replicate.
  • Inference may exceed training. Per-token economics scale with adoption.
  • Enterprises and governments are early. Sovereign AI orders add a new buyer category.
  • Hyperscaler capex remains elevated. Microsoft, Google, Amazon, Meta, Oracle are still committing capital.
  • Ecosystem investments accelerate adoption. Nvidia-backed deals expand TAM faster than organic growth.
  • Software and systems mix. Revenue per system rises as more of the stack is sold by Nvidia.
Bear case

Why the cycle could weaken

  • Customer concentration. Four direct customers carry >10% revenue each.3
  • Circular financing. Supplier-backed deals may flatter independent demand.46
  • Neocloud balance-sheet stress. Rented GPU economics depend on utilisation and pricing.
  • Hyperscaler ASICs. Internal Trainium / TPU / Maia / Axion erodes Nvidia’s share at the margin.
  • AMD and custom accelerators. Competition compresses pricing power over time.
  • Model efficiency. Quantisation, sparsity, and better inference reduce compute intensity per token.
  • China and export controls. Address-ability shrinks structurally.2
  • Overbuilding. Capacity may run ahead of the AI revenue that justifies it.

The strongest bull case is that today’s capex looks aggressive only because the AI economy is still young. The strongest bear case is not that AI is fake, but that infrastructure is being built faster than the revenue models that justify it.


XI. The composite view

Both cases can be partly right at once. The bull case can carry the next few quarters while the bear case slowly accumulates pressure on the back end of the cycle. The honest 2026 reading is that Nvidia’s earnings are real and large, AI capex is real and durable in aggregate, but the quality of those earnings will be tested as the next wave of installed capacity has to convert into utilisation and cash flow.

The earnings are real. The question is whether the demand behind them is repeatable without supplier-backed financing forever.


XII. What to watch

This is a working checklist, not a prediction. A few signals will move first if the bear case starts to bite, and others will reinforce the bull case if AI usage actually broadens.

What to watch
  • Data Center revenue growth versus customer concentration.
  • Gross margin as Blackwell / Rubin full systems ramp.
  • China revenue and export-control changes.
  • Customer prepayments, commitments, and financing structures.
  • Neocloud utilisation and GPU rental pricing.
  • Hyperscaler capex versus their own AI revenue.
  • Nvidia investments in customers and partners.
  • HBM supply and pricing.
  • CoWoS / advanced packaging capacity.
  • Inference revenue growth across the ecosystem.
  • Internal ASIC adoption by hyperscalers.
  • Depreciation burden across cloud customers.
  • Enterprise AI adoption and pricing.
  • Sovereign AI orders.

Quick terms


XIII. The earnings quality test

In 2021, Nvidia’s earnings-quality question was crypto. The risk was that mining demand made temporary GPU demand look like permanent gaming growth.

In 2026, the question is AI capex. The risk is not that Nvidia’s revenue is fake. The risk is that some part of the demand may be pulled forward by customers racing to build capacity before the economics are fully proven.

The revenue is real. The demand is real. The AI infrastructure buildout is real. But the final test is utilisation. Do the GPUs get used. Do the AI products generate cash. Do the customers earn returns. Do they come back for more without needing supplier financing. Do the economics survive model efficiency, custom chips, export controls, and power constraints.

The earnings are real. The question is who ultimately pays for the AI buildout.

That is Nvidia’s earnings quality test.


1 Patel, D. (May 2021). Nvidia is Raking in the Money, but are Earnings Inflated? SemiAnalysis. Historical anchor for the 2021 crypto / CMP earnings-quality framing. Used as inspiration only. No content, structure, or charts reproduced.

2 Nvidia (FY2026). Q4 and FY2026 financial results. Q4 FY2026 revenue $68.1B, full-year FY2026 revenue $215.9B, Q4 Data Center revenue $62.3B, and commentary on the Hopper-to-Blackwell transition, the H20 charge, and forward outlook assumptions excluding Data Center compute revenue from China. Used as confirmatory context for margins and the China exposure paragraph.

3 Nvidia (Q3 FY2026). Form 10-Q. Four direct customers each above 10% of total revenue, with shares of approximately 22%, 15%, 13%, and 11%, primarily attributable to Compute & Networking, alongside concentration language across direct, indirect, and cloud-service purchasers.

4 Nvidia (2025). Nvidia and OpenAI announce strategic partnership. At least 10 GW of AI data centers using Nvidia systems, Nvidia intent to invest up to $100B in OpenAI progressively as each gigawatt is deployed, with the first gigawatt planned for the second half of 2026 on Vera Rubin.

5 OpenAI (2025). OpenAI & Nvidia systems partnership. Counterpart statement of the same partnership used for confirmatory context.

6 Reuters (2025-26). Coverage of the Nvidia / CoreWeave relationship, including reporting on a roughly $2B Nvidia investment in CoreWeave and a multi-year agreement under which Nvidia agreed to purchase unsold cloud capacity for up to $6.3B through 2032, with investor commentary raising questions about circular financing. Cited via Reuters distribution.

7 TSMC. 2025 Annual Report. AI/HPC as a structural driver of advanced process demand and continued packaging investment, framed alongside mild non-AI recovery.

8 ASML (2025). 2025 Annual Report, strategic report section. AI requires leading-edge, high-performance processor chips and a significant increase in DRAM compared with traditional compute architectures.

Further reading
*   *   *

This is Essay No. 020. The topics: intelligence, AI, systems, knowledge, and the questions underneath the questions everyone else is asking. If you read this far and disagreed with any part of it, write to me. I read everything.

Pugalenthi Magendran