← Back to blog
Essay No. 033  ·  AI Infrastructure  ·  Melbourne, Australia
AI Infrastructure Nvidia AI Factory CUDA Blackwell Rubin TSMC HBM CoWoS Datacenter Hyperscalers GPU ASML Capex

The Bubble That Became Infrastructure.Original analysisNot investment advice

Why Nvidia’s 2021 overvaluation story turned into the AI factory thesis.
PM
Pugalenthi Magendran
April 2026  ·  Melbourne, Australia
12 min read

The 2021 Nvidia bubble thesis was right about crypto, gaming, stock-split speculation, and datacenter digestion. But it underestimated the AI factory. Nvidia did not stay a GPU company waiting for the bubble to pop. It became the operating layer of accelerated computing. In 2026, the question is not whether Nvidia is real. It is whether the AI infrastructure cycle can keep producing enough economic value to support the scale Nvidia has reached.

In 2021, SemiAnalysis called Nvidia overvalued. Not bad. Overvalued. That distinction matters.

The uploaded article actually praised Nvidia heavily. It said Nvidia was the backbone of the AI revolution, the semiconductor company that understood software, and one of the few companies prepared for the AI and data-centric world. But then it made the financial argument: the stock price had run ahead of the business.1

The article blamed a familiar set of forces. Crypto. Gaming. COVID pull-forward. Datacenter buildout. Mellanox acquisition math. Stock-split speculation. That was the 2021 bubble thesis.

In 2026, the story looks different. Crypto demand faded. Ethereum mining ended. Gaming normalised. Stock splits still did not create enterprise value. But Nvidia did not return to being a normal GPU company. The opposite happened.

AI turned Nvidia into infrastructure.

Key idea

The 2021 Nvidia bubble thesis was right about temporary demand and speculative market behavior, but wrong about the size of the coming AI platform shift. Nvidia’s crypto/gaming bubble did not become the future. Its AI factory platform did. In 2026, the risk is no longer a simple stock-split bubble. The risk is whether AI infrastructure spending can generate enough real economic returns to support Nvidia’s revenue scale, margins, customer concentration, and supply-chain commitments.


I. The 2021 thesis was not stupid

In July 2021, Dylan Patel published a SemiAnalysis piece that praised Nvidia’s vision and software understanding while arguing the stock had run ahead of fundamentals. The piece walked through the 4-for-1 stock split (with shares rising roughly 40% after the split announcement), compared the 2021 setup with Nvidia’s 2018 bubble, and laid out a set of temporary or cyclical drivers: crypto GPU mining demand, COVID gaming pull-forward, datacenter buildout and digestion, the Mellanox acquisition flattering year-over-year datacenter growth, and Tesla-style stock-split speculation behavior. The deeper conclusion was that Nvidia was a great company, but temporary growth was being extrapolated too aggressively.1

Five years on, that combination of praise and caution looks intellectually honest. The cyclical reads were correct. The platform read was incomplete.

2021 thesis

Nvidia was a great company, but the market was extrapolating temporary crypto, gaming, datacenter, and stock-split-driven momentum too aggressively.

Diagram · The 2021 bubble ingredients
Temporary 01

Crypto GPUs

Ethereum proof-of-work mining demand for consumer GPUs.1
Temporary 02

COVID gaming

Pull-forward of gaming GPU demand during lockdowns.1
Cyclical 03

DC buildout

Datacenter buildout phase ahead of digestion.1
Speculative 04

Stock split

4-for-1 split adding ~40% post-announcement momentum.1
A simplified, original visual of the four bubble ingredients identified by the 2021 SemiAnalysis piece.

II. Crypto was temporary

The Ethereum Foundation describes The Merge as completing on September 15, 2022, moving the network from proof-of-work to proof-of-stake.4 That single event removed the most important GPU-mining demand source. Not all crypto disappeared, but the GPU-mining tailwind that had inflated 2021 expectations did fade.

Crypto was not the future of Nvidia demand. It was noise around the transition to something much bigger.


III. Gaming normalised, but Nvidia did not

The 2021 piece worried about gaming digestion after COVID, and that worry was directionally correct: gaming growth above the COVID line was not sustainable.1 But the right way to read the company in 2026 is mix. Gaming remained important, but the centre of gravity moved to Data Center.2

Gaming did not have to collapse for the 2021 thesis to age badly. Data Center simply became much larger.


IV. The business caught up to the imagination

Nvidia’s FY2026 results page reports full-year revenue of about $215.9B, Q4 FY2026 revenue of about $68.1B, Q4 Data Center revenue of about $62.3B, and full-year Data Center revenue of about $193.7B.2 In 2021, Nvidia’s valuation looked detached from fundamentals. By 2026, the revenue base had changed completely.

Card · Nvidia FY2026 results, simplified
~$215.9B
FY2026 revenue2
~$193.7B
FY2026 Data Center revenue2
~$68.1B
Q4 FY2026 revenue2
~$62.3B
Q4 FY2026 Data Center revenue2
Figures as reported by Nvidia in its FY2026 results page. Treated as company disclosures; this essay is not investment advice.

Nvidia did not become cheap. The business became enormous.


V. The GPU became a system

Nvidia’s FY2026 Form 10-K describes the company as a data-center-scale AI infrastructure platform. The filing details a technology stack that pairs CUDA with hundreds of domain-specific software libraries, frameworks, algorithms, SDKs, and APIs, and frames Blackwell as data-center-scale infrastructure spanning GPUs, CPUs, DPUs, interconnects, switch chips, systems, and networking adapters — with hundreds of thousands of GPUs interconnected to function as a single giant computer. The filing also positions Rubin for production shipments in the second half of FY2027 and frames the next generation around agentic AI, reasoning, and long-context workflows.3

Diagram · 2021 GPU company vs 2026 AI factory company
2021 framing

GPU company

GPU silicon
Board / module
Server
Customer integrates
2026 framing

AI factory company

GPU + CPU + DPU silicon
NVLink + InfiniBand / Spectrum-X
Rack-scale systems
CUDA + libraries + SDKs
Cloud + enterprise + sovereign
Inference + serving software
AI factory architecture
A simplified, original split. The product surface area widened from a chip to a factory architecture.3

Nvidia’s moat is not the GPU alone. It is the software-defined system around the GPU.


VI. CUDA became the operating layer

CUDA is not only syntax. It is developer habit, libraries, kernels, training scripts, inference tooling, profiling, debugging, cloud images, ML frameworks, hiring market, Stack Overflow answers, research defaults, enterprise support, and benchmark recipes. Nvidia’s 10-K cites CUDA and hundreds of software libraries, frameworks, algorithms, SDKs, and APIs as part of the platform.3

Diagram · Nvidia full-stack AI factory, simplified
01
Silicon GPU, CPU, DPU, NVLink switch, NIC
Hardware
02
Systems Blackwell rack, networking, cooling
System
03
Software CUDA + hundreds of libraries / SDKs / APIs3
Software
04
Workloads training, inference, agentic AI, reasoning, physical AI3
Workloads
05
Customers cloud, enterprise, sovereign AI
Demand
A simplified, original visual based on Nvidia’s public framing in its FY2026 10-K. Not an Nvidia chart.3

A competitor can build a chip. Nvidia sells the default path.


VII. The supply chain became the product

Nvidia’s FY2026 10-K describes a supply chain that spans foundries such as TSMC and Samsung, memory partners including SK hynix, Micron, and Samsung, advanced packaging such as CoWoS, and contract manufacturers including Hon Hai, Wistron, and Fabrinet for assembly, testing, packaging, and final products.3 ASML’s 2025 Annual Report frames AI as requiring leading-edge high-performance processors and a significant increase in DRAM relative to traditional compute architectures.5 TSMC’s 2025 Annual Report and 2026 symposium materials describe robust AI demand and ongoing CoWoS expansion for compute-and-memory integration.67

Diagram · The supply chain behind every Nvidia sale
Wafers
TSMC / Samsung
advanced logic node3
Memory
SK hynix / Micron
HBM stacks3
Packaging
CoWoS
compute + HBM integration7
Networking
NVLink / InfiniBand
scale-up + scale-out
Systems
ODMs
rack assembly + test3
Site
Power + DC
cooling, energy, capacity
A simplified, original map of the physical stack pulled by every AI accelerator sale. Not exhaustive; not an Nvidia or supplier chart.

AI demand became semiconductor demand. Semiconductor demand became Nvidia revenue.


VIII. The new bubble question is capex digestion

The 2021 bubble question was whether Nvidia was inflated by crypto, gaming, COVID, and a stock split. The 2026 bubble question is whether AI factories are being built faster than AI economics can justify. Nvidia’s own 10-K is explicit about the demand-forecasting risk: if Nvidia underestimates demand and suppliers cannot increase production, it may lose revenue and market share; if Nvidia overestimates demand or customers cancel, defer, or buy competitors, Nvidia may face excess inventory, purchase commitments, pricing pressure, and margin pressure.3

Reading the cycle

The digestion risk did not disappear. It scaled.

Hyperscaler capex commentary across Alphabet, Meta, Microsoft, and Amazon points to large multi-year AI infrastructure spend through 2026 and beyond.8910 Each of those investments has to produce enough AI revenue or cost savings to justify itself. The new bubble question is not whether AI is fake; it is whether the build pace matches the monetisation pace.

Diagram · The AI capex digestion loop
01
Capex
hyperscaler spend
02
Buildout
Nvidia + ODM + DC
03
Utilisation
training + inference
04
Revenue
AI products + APIs
05
Next capex
expand or pause
← the loop closes; pace must match monetisation
A simplified, original 5-step loop. The thesis is not that the loop is broken; it is that each step has to keep working for the next capex cycle to repeat.

The new bubble question is not whether AI is fake. It is whether AI infrastructure is being built faster than AI revenue can justify.


IX. Customer concentration matters

Nvidia’s FY2026 10-K reports that one direct customer represented 22% of total revenue and another represented 14%, both primarily attributable to Compute & Networking. The filing also notes that one AI research and deployment company is estimated to have contributed meaningful revenue through cloud services purchased from Nvidia customers.3

This is not automatically bearish. Large infrastructure markets often begin concentrated. But it means demand depends heavily on a small number of massive buyers.

Nvidia’s growth is massive, but the buyers are massive too.


X. Regulation is now structural

Nvidia’s FY2026 10-K discusses export-control risks, geopolitical risk around China, Taiwan, and Korea, regulatory inquiries and competition-law scrutiny, and the risk that some customers may build or use their own solutions.3 Nvidia is no longer just a chip company; it is strategic infrastructure.

Diagram · The structural risk surface
Export controls

Reduced accessible markets

Restrictions on advanced GPU exports to specific countries and end users reduce the addressable customer base and complicate product roadmaps.3
Competition / antitrust

Bundling and allocation scrutiny

Regulators may scrutinise bundling, supply allocation, and customer arrangements as Nvidia’s share of accelerated compute grows.3
Geopolitics

Taiwan, Korea, China concentration

Supply concentration in Taiwan and Korea exposes Nvidia to geopolitical and trade-policy shocks.3
Customer in-house silicon

Customers may build their own

Hyperscalers and some enterprises have the capability to build or use their own AI silicon.3
A simplified, original visual of the structural risks Nvidia’s own filings highlight. Risk dimensions, not predictions.

Nvidia became too important to be treated like a normal chip stock.


XI. Custom silicon is the real competitive threat

Nvidia’s 10-K notes that some customers have internal development capabilities and may build or use their own solutions.3 The visible examples are well known. The largest Nvidia customers are also the most capable potential competitors.

Diagram · Customer custom silicon, simplified
Hyperscaler
Google TPU
internal AI accelerator family
Hyperscaler
AWS Trainium / Inferentia
AWS-designed AI silicon
Hyperscaler
Microsoft Maia
Microsoft custom AI chip
Application
Meta MTIA
Meta custom AI silicon family
ASIC partner
Broadcom
custom AI ASICs for hyperscalers
Ecosystem
NVLink Fusion
heterogeneous integration around NVLink
A simplified, original map of the custom-silicon surface around Nvidia’s largest customers. Not exhaustive; not a market-share claim.

The largest customers are also the most capable future competitors.


XII. The 2021 thesis aged half right

A clean way to read the 2021 thesis is to score it on what aged right versus what aged wrong.

Diagram · The 2021 thesis — what aged right vs what aged wrong
Aged right

Temporary forces faded

  • Crypto. Ethereum’s 2022 Merge ended the GPU-mining tailwind.4
  • Stock split. A split does not create enterprise value.1
  • Gaming. Growth above the COVID line was not sustainable.1
  • Digestion. Datacenter cycles still have digestion phases.3
  • Extrapolation. Temporary growth should not be extrapolated forever.
Aged wrong

Platform shift was larger

  • Software moat. CUDA + libraries compounded into the operating layer.3
  • Networking. Mellanox / NVLink / Spectrum-X became core product.3
  • Data Center. The segment became the company.2
  • System scope. Nvidia moved from GPU vendor to AI infrastructure company.3
  • AI capex. Industrial-scale, multi-year hyperscaler buildouts.89
A simplified, original split. The article was right about the bubble ingredients. It was incomplete about the platform outcome.

The article was right about the bubble ingredients. It was wrong about the platform outcome.


XIII. What could break the thesis?

The strongest bear case is not that Nvidia is fake. It is that AI capex may be cyclically overbuilt.

Bear case · what could break the thesis
  1. Overbuild. AI infrastructure buyers overbuild relative to monetisation.
  2. Capex slowdown. Hyperscalers slow buildouts after a digestion period.8
  3. Slow absorption. AI revenue does not absorb infrastructure investment fast enough.
  4. Inference margin compression. Inference unit economics tighten as competition rises.
  5. Custom silicon share. Hyperscaler and enterprise ASICs reduce Nvidia dependence.3
  6. AMD credibility. A stronger AMD MI roadmap absorbs some of the alternative slot.
  7. Export controls. Restrictions reduce accessible markets.3
  8. Regulatory pressure. Scrutiny of bundling, allocation, or customer agreements.3
  9. Supply bottlenecks. HBM, CoWoS, substrates, power, and rack-scale deployment cap growth.7
  10. Product transitions. Blackwell-to-Rubin transition creates inventory or margin risk.3
  11. Concentration. A pause by major buyers shows up quickly in revenue.3
  12. Power and capacity. Energy and data-center capacity constrain growth.
  13. Supplier scale. Nvidia’s own suppliers cannot scale fast enough.

XIV. What could break the bear case?

The strongest bull case is that Nvidia is not selling one chip cycle. It is selling the operating layer of accelerated computing.

Bull case · what could break the bear
  1. AI is the new workload. Models and agents become embedded across software.
  2. Inference scale. Inference demand may exceed training demand by a large margin.3
  3. Agentic AI. Multi-step reasoning and tool use raise token consumption.3
  4. Physical AI. Robotics and embodied AI add a new demand axis.3
  5. Enterprise AI. Enterprise adoption is still early.
  6. CUDA moat. The software moat keeps compounding.3
  7. Cadence. Blackwell / Rubin keep performance ahead of competitors.3
  8. System lock-in. Rack-scale systems and networking raise switching costs.3
  9. Mixed buyers. Hyperscalers can buy Nvidia and build custom silicon.
  10. Long-life infra. AI factories may become long-life infrastructure like cloud regions.

Nvidia is not selling one chip cycle. It is selling the operating layer of accelerated computing.


XV. What to watch

What to watch
  • Data Center revenue growth.2
  • Gross margin trend.
  • Blackwell Ultra ramp.3
  • Rubin shipment timing.3
  • NVLink Fusion adoption.
  • HBM supply trajectory.5
  • CoWoS capacity expansion.7
  • Networking revenue mix.
  • Customer concentration.3
  • Hyperscaler capex commentary.89
  • AI revenue at hyperscalers.
  • Inference monetisation evidence.
  • Cloud GPU utilisation.
  • Power and DC bottlenecks.
  • Export-control updates.3
  • Regulatory investigations.3
  • Custom silicon adoption.
  • AMD MI-series traction.
  • Inventory and purchase commitments.3
  • Product-transition risks.
  • AI factory ROI evidence.

Glossary

A short reference for the vocabulary used above. Definitions are simplified.

Glossary
GPU
Graphics processing unit, now widely used for parallel AI computation.
CUDA
Nvidia’s programming platform for GPU computing.
Data Center revenue
Nvidia segment covering accelerated compute, networking, AI solutions, and related software.
AI factory
Data-center infrastructure optimised to produce AI training and inference outputs.
HBM
High-bandwidth memory used near AI accelerators.
CoWoS
TSMC advanced packaging used to integrate compute and HBM.
NVLink
Nvidia high-speed interconnect.
DPU
Data processing unit for networking, storage, and infrastructure offload.
Inference
Running trained models to generate outputs.
Training
Process of creating or updating model weights.
Capex
Capital expenditure used to build infrastructure.
Datacenter digestion
Period where customers absorb already-purchased infrastructure before buying more.
Customer concentration
Revenue dependence on a small number of large buyers.
Export controls
Government restrictions on shipping advanced chips to certain countries or end users.
Custom silicon
Chips designed by cloud companies or customers for their own workloads.

XVI. The bubble that became infrastructure

The 2021 Nvidia bubble thesis was right about crypto, gaming, stock-split speculation, and datacenter digestion.

But it underestimated the AI factory.

Nvidia did not stay a GPU company waiting for the bubble to pop. It became the operating layer of accelerated computing. The chip became a system. The system became a platform. The platform became infrastructure.

In 2026, the question is not whether Nvidia is real. It is whether the AI infrastructure cycle can keep producing enough economic value to support the scale Nvidia has reached.

Reasonable critics in 2021 looked at speculative behaviour and called it speculative behaviour. That was correct. They underestimated the size of the platform shift that arrived under it. AI absorbed the speculative energy and turned it into industrial-scale demand.

Reasonable critics in 2026 should look at AI capex and ask the right question now. Not whether AI works. Whether the buildout pace matches monetisation.

That is the bubble that became infrastructure.


1 Patel, D. (Jul 2021). Nvidia’s 2021 Bubble, Eerily Similar To Other Bubbles That Came Before. SemiAnalysis. Historical anchor for the 2021 bubble framing, including praise for Nvidia’s vision and software understanding, the overvaluation argument, the 4-for-1 stock split and ~40% post-announcement run, the 2018 comparison, datacenter buildout / digestion, Mellanox acquisition math, crypto GPU mining, Ethereum proof-of-stake expectation, COVID gaming pull-forward, Tesla stock-split comparison, and the deeper point that temporary growth was being extrapolated too aggressively. Used as inspiration only. No content, structure, or charts reproduced.

2 NVIDIA (2026). NVIDIA announces financial results for fourth quarter and fiscal 2026. FY2026 revenue, Q4 FY2026 revenue, full-year Data Center revenue, and Q4 Data Center revenue framing used in this essay. Reported as company disclosures.

3 NVIDIA (2026). Form 10-K, fiscal year 2026. Data-center-scale AI infrastructure framing, CUDA + hundreds of software libraries and frameworks, Blackwell as a full system (GPU, CPU, DPU, interconnects, switches, networking), hundreds of thousands of GPUs interconnected as a single computer, Rubin production timing in H2 FY2027 with agentic AI / reasoning / long-context positioning, supply-chain disclosures (TSMC, Samsung, SK hynix, Micron, CoWoS, Hon Hai, Wistron, Fabrinet), demand-forecasting risk, customer concentration (22% and 14% direct customers), export controls, geopolitics, regulatory inquiries, and customer in-house silicon language.

4 Ethereum Foundation. The Merge. Completed September 15, 2022; transition from proof-of-work to proof-of-stake.

5 ASML (2025). 2025 Annual Report, strategic report section. AI requires leading-edge high-performance processors and a significant increase in DRAM relative to traditional compute architectures.

6 TSMC. 2025 Annual Report. Robust AI-related demand and the role of advanced logic and packaging for AI/HPC.

7 TSMC (2026). 2026 North America Technology Symposium. CoWoS expansion for compute-and-memory integration, advanced packaging roadmap, AI/HPC packaging direction. Specific reticle-scaling and HBM-stack claims used here only at the level the cited material supports.

8 Alphabet (2026). Alphabet investor relations. Q1 2026 capex and technical-infrastructure / AI spending commentary used as context for hyperscaler AI capex digestion risk.

9 Meta (2026). Meta investor relations. 2026 capex guidance and AI infrastructure / data-centre spending commentary used as context for hyperscaler AI capex digestion risk.

10 Microsoft / Amazon (2026). Microsoft IR and Amazon IR. Cloud AI capex commentary and AI infrastructure demand framing used at the level the cited investor materials support.

11 NVIDIA. Public GTC 2026 / Rubin / Blackwell Ultra materials are referenced in this essay at the level Nvidia’s own SEC filing already discloses. Treated as Nvidia statements rather than independent verification.

Further reading
*   *   *

This is Essay No. 033. 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