📈
Related Insight
AI Stocks: Who Actually Won Since ChatGPT?
NVDA +800% (2022 trough to 2024 peak). SK Hynix +2,000%. The four waves — and what comes next for your portfolio.
→
Insights — Q2 2026

The AI Spending Bubble:
Who's Actually Making Money vs. Burning It

The AI industry is simultaneously the most profitable hardware cycle in semiconductor history and the largest single-year loss-making experiment in corporate history — depending on where you look. NVIDIA made $120 billion in GAAP net income last fiscal year ($117B non-GAAP). OpenAI is projected to lose more than $14 billion in 2026. Forty percent of AI startups launched in 2024 are already closed. We followed every dollar of AI spending across five layers of the stack — and the picture that emerges is stranger, and more precarious, than the headlines suggest.

Last updated June 9, 2026 · ~13 min read · Sources: NVIDIA, OpenAI, Goldman Sachs, MIT, CB Insights, Morgan Stanley
$120B
NVIDIA Non-GAAP Net Income FY2026
71% full-year gross margins (75% in Q4). Most profitable hardware cycle in history.
$14B+
OpenAI Projected Loss · 2026
$13B revenue vs. ~$60B compute costs annually.
40%
AI Startups Closed in 24 Months
14,000+ launched in 2024. 5,600+ shut by mid-2026.
$700B
Hyperscaler Capex Pledged · 2026
Nearly double 2025. Amazon alone projects $200B.
The Full Picture

The Money Map — Five Layers, Five Very Different Stories

AI spending flows through five distinct layers, and the profit-and-loss picture at each layer is radically different. Click each layer to see the numbers.

Interactive · Click each layer to expand
Where the AI Money Goes — and Who Keeps It
Five layers of the AI stack. Profits flow upward to infrastructure. Losses concentrate in the middle. The application layer is a graveyard with survivors.
$216B
NVIDIA Revenue FY2026
75%
NVIDIA Gross Margin
$120B
NVIDIA GAAP Net Income FY2026
92%
NVIDIA GPU Market Share
This is the single most profitable layer in the entire AI stack — and it's not close. NVIDIA's FY2026 revenue of $215.9 billion was up 65% year-over-year. Data center revenue alone was $197.3 billion — 91% of the total. GAAP net income of $120 billion (non-GAAP: $117B) at 71.1% full-year gross margins (75% in Q4 alone) makes NVIDIA arguably the highest-margin large industrial company in history. For every dollar a hyperscaler spends on AI training, NVIDIA captures approximately 55–65 cents in gross profit.1
NVIDIA (NVDA)
$120B net income
GAAP; non-GAAP $117B. 71% full-year gross margin (75% Q4); $4.5T market cap
TSMC (TSM)
~$35B net income
Sole manufacturer of AI chips at N3/N2; backlog through 2027
SK Hynix
HBM monopoly
Primary HBM3E supplier for H200/B100; demand 2–3× by 2027
Broadcom (AVGO)
~+730% since 2022
Custom AI ASICs for Google, Meta; networking for AI clusters
AMD
~+80% since 2022
MI300X gaining share, but distant #2; CUDA moat challenge
+508%
Vistra (VST) Since Nov 2022
+300%
Constellation (CEG) Since Nov 2022
$28B
CoreWeave Debt Raised
2×
US Data Center Power Demand by 2028
The power layer is a legitimate structural winner: data center electricity demand in the US is projected to nearly double from 80 to 150 gigawatts between 2025 and 2028, with AI the primary driver. Nuclear energy operators with 24/7 clean power — essential for data center sustainability commitments — have been re-rated accordingly. Vistra and Constellation have signed multi-year fixed-price power purchase agreements at premiums that would have been unthinkable in 2021. The one wildcard: CoreWeave, the GPU-as-a-service company, has raised $28 billion in debt backed primarily by Microsoft and Meta contracts — its creditworthiness is borrowed, not owned.2
$700B
Combined Capex Pledged · 2026
34%
AI Capex-to-Sales Ratio · 2026E
+75%
Azure AI Revenue Growth
$1T
Undisclosed Lease Commitments
The hyperscalers are not in crisis — but they are making a bet of extraordinary scale. Their AI capex-to-sales ratio is projected to reach 34% in 2026, approaching the 32% dot-com peak in 2000. Morgan Stanley's Todd Castagno has noted this directly. The difference from 2000: these companies are funding capex entirely from free cash flow generated by highly profitable existing businesses. Microsoft, Amazon, Alphabet, and Meta collectively generated over $350 billion in operating cash flow in 2025 — the AI build-out is affordable, but not trivially so. Oracle is a more fragile story: capex has been 66% of revenues in early FY2026, and its stock has fallen over 57% since pledging $300 billion in AI infrastructure for OpenAI in September 2025.3
Microsoft
Azure AI growing
$120B+ capex in 2026; OpenAI dependency is 2-way
Amazon (AWS)
~$200B capex 2026
Bedrock AI platform growing; largest single hyperscaler spend
Alphabet
~$180B capex 2026
TPU advantage; issued 100-year bond to finance infrastructure
Meta
~$125B capex 2026
No cloud business; AI embedded in ads engine directly
Oracle
Capex 66% of revenue
Stock −57% since Stargate announcement; single-digit cloud margins
$14B+
OpenAI Projected Loss · 2026
$44B
OpenAI Projected Cumulative Losses to 2028
$13B
OpenAI Revenue · 2026E
$730B
OpenAI Valuation
The foundation model layer is where the most extraordinary financial tension in the AI economy lives. OpenAI's valuation grew from $80 billion to $730 billion in under three years — while projecting cumulative operating losses of $44 billion through 2028. Revenue of approximately $13 billion in 2026 is real but covers only a fraction of compute costs estimated at ~$60 billion per year. The company has committed to spending $665 billion by 2030, targets cash-flow positivity in 2030, and in January 2026 it was reported to have missed its own internal revenue targets for 2025. OpenAI's relationship with Oracle is particularly revealing: Columbia Business School professor Stijn Van Nieuwerburgh described it as "renting Oracle's creditworthiness" — OpenAI needs Oracle's investment-grade rating to finance data centers it cannot afford on its own balance sheet.4
OpenAI
$14B+ loss · 2026E
$730B valuation; $44B projected losses through 2028
Anthropic
Unprofitable
$380B valuation; potential 2026 IPO for clarity on timeline
xAI (Elon Musk)
Burning
Relying on Tesla and X infrastructure; no independent financials disclosed
DeepSeek
$6M training run
Matched OpenAI o1 at 1/16th the cost — shook the entire layer
40%
Failure Rate in 24 Months
95%
AI Pilots Failing to Scale (MIT · 2025)
25%
AI Initiatives Delivering Expected ROI (IBM)
498
AI Unicorns Still Standing
This is where the AI bubble is most clearly visible. Over 14,000 AI startups launched globally in 2024 alone. By mid-2026, approximately 40% had closed — 3,800 in 2025, another 1,800 in early 2026. The causes are consistent: OpenAI's product cadence directly cannibalized at least 200 funded startups in 2024 by bundling their core features into existing products; GPU compute costs of $1M+ per month combined with declining willingness-to-pay as free alternatives proliferated; and most critically, startups built without proprietary data — "wrappers" around foundation model APIs with no moat, one product update away from obsolescence. The survivors cluster in regulated verticals where foundation model companies won't compete: healthcare, legal, finance, and compliance infrastructure.5
✦ The Core Tension in One Sentence
The AI economy currently works as follows: NVIDIA extracts ~60 cents of gross profit for every dollar hyperscalers spend; hyperscalers spend those dollars to build infrastructure they hope will generate cloud AI revenue; that cloud AI revenue flows partly to foundation model companies that are burning it faster than it arrives; and the application layer is subsidized by investor capital that is running out of patience. The entire stack is in equilibrium only as long as hyperscaler capex keeps growing.
Layer 1 & 2 Deep Dive

The Clear Winners: Infrastructure Is Printing Money

The single most reliable way to profit from an AI boom — historically and currently — is to sell the tools and infrastructure, not the intelligence itself. The 2020s version of this is NVIDIA. And the margins are extraordinary.

NVIDIA — The Clearest Winner
$81.6B quarterly revenue, $58.3B quarterly GAAP net income: the most profitable quarter in semiconductor history
NVIDIA's Q1 FY2027 results (Feb–April 2026) broke every record: $81.6 billion in revenue, up 85% year-on-year, with $58.3 billion in GAAP net income (non-GAAP net income: $45.5B) — a 37% increase from the prior quarter and more than 200% year-on-year. The company forecasts $91 billion in revenue for Q2. Its order book through 2027 is reported to exceed $1 trillion. CUDA — its programming framework — has been adopted by AI researchers for over a decade, making the software moat as durable as the hardware advantage. The single risk: DeepSeek demonstrated in January 2026 that training frontier models at 1/16th the compute cost is possible — a development that briefly erased $600 billion from NVIDIA's market cap in a single day.
Energy Infrastructure — The Second Wave Winner
Nuclear energy repriced by data center demand. Vistra: +508%. Constellation: +300%.
AI data centers have a unique power requirement: 24/7 guaranteed clean electricity — something that only nuclear can reliably deliver at scale. This created a structural re-rating of nuclear power operators that most investors had left for dead a decade earlier. Vistra's +508% return and Constellation's +300% return since ChatGPT's launch in November 2022 reflect fixed-price, long-term power purchase agreements at premiums that would have been unthinkable when these plants were considered stranded assets. The US data center electricity demand is expected to nearly double from 80 to 150 GW between 2025 and 2028, with no sign of slowdown.
The Uncomfortable Truth About Infrastructure Profits
The AI economy's profits are concentrated at the picks-and-shovels layer — the same companies that will survive if the application layer implodes
NVIDIA makes money whether OpenAI is profitable or not. TSMC gets paid whether the hyperscalers' AI cloud revenue justifies their capex or not. Nuclear energy operators have locked in 15–20 year power purchase agreements that will pay out regardless of whether AI achieves its productivity promise. This is the structural asymmetry of the current AI economy: the infrastructure layer has locked-in revenue from a counterparty (hyperscalers) that can fund it from cash flow; the application layer's revenue depends on enterprises actually finding value — which, per MIT, is failing in 95% of cases. If the application layer disappoints at scale, infrastructure doesn't immediately feel it. If the hyperscalers slow capex, infrastructure feels it very quickly.
Layer 3 Deep Dive

The Subsidisers: Hyperscalers Are Spending Before They Earn

Microsoft, Amazon, Alphabet, and Meta are not in financial distress — but they are making spending commitments at a scale and pace that requires extraordinary revenue growth to justify. The math is only comfortable if AI cloud adoption accelerates sharply over the next three years.

34%
Capex-to-Sales Ratio · 2026 Projected
AI capex as a share of revenue has reached dot-com era levels — with a crucial difference
Morgan Stanley's analysis puts hyperscaler AI capex-to-sales at 34% in 2026 — approaching the 32% dot-com peak in 2000, and projected to reach 37% by 2028. The crucial difference from 2000: the companies doing the spending (Microsoft, Amazon, Alphabet, Meta) are the four most profitable companies in history by operating cash flow, generating over $350 billion in operating cash flow collectively in 2025. The telecom companies that overbuilt in 1999–2001 were largely financed by debt. The hyperscalers are largely financing from free cash flow. But "financed from cash flow" does not mean "the returns are certain."
34% Capex/Revenue 2026E 37% Projected by 2028
$20
The Subsidy Problem · Microsoft & GitHub Copilot
Microsoft was losing $20 per user per month on GitHub Copilot. It has just switched to usage-based pricing.
The clearest sign that AI product economics are changing: GitHub Copilot shifted from subscription to usage-based pricing in June 2026, after Microsoft was reported to be losing approximately $20 per user per month on the $10/month subscription. The era of subsidized AI pricing — where compute costs were deliberately obscured behind flat-rate subscriptions to drive adoption — is ending. When a single premium request can consume $11 in compute, the flat-rate model collapses. ServiceNow and Uber reportedly burned through their entire annual AI token budgets before mid-year. Usage-based pricing will force a genuine reckoning with AI ROI at the enterprise level.
$20/user/month loss on Copilot Usage pricing begins June 2026
$1T
The Off-Balance-Sheet Risk
Hyperscalers have nearly $1 trillion in undisclosed future lease commitments for data centers not yet built
Amazon, Meta, Alphabet, Microsoft, and Oracle have accumulated close to $1 trillion in off-balance-sheet future lease commitments for data centers that haven't been built yet. These are contractual obligations to pay for infrastructure regardless of whether the AI revenue materialises to cover it. Alphabet has issued long-duration bonds to fund infrastructure. Oracle took on $43 billion in debt in FY2026 alone. These are rational financing decisions if AI revenue scales as projected — but they represent a significant claims on future earnings that most casual analysis of hyperscaler "strength" doesn't fully account for.
~$1T in undisclosed lease obligations
✦ Why the Hyperscalers Are Still the Safest Bet in This Stack
Even if AI application revenue disappoints, the hyperscalers have a fallback: cloud computing itself. AWS, Azure, and Google Cloud generate combined annual revenues exceeding $400 billion from non-AI workloads. The AI build-out is an option on additional revenue — the downside is lower growth, not insolvency. This is fundamentally different from the foundation model companies, whose entire revenue thesis depends on AI monetisation working.
Layer 4 Deep Dive

The Burners: Foundation Models and the $44 Billion Bet

OpenAI, Anthropic, and the other foundation model labs are making the largest single bet in corporate history: that intelligence, once commoditised at scale, will be worth more than the extraordinary costs to build it. The numbers required to justify the bet are staggering.

Company Valuation Revenue ~2026E Loss / Status Path to Profit
OpenAI $730B ~$13B $14B+ loss Cash-flow positive target: 2030. Requires ~$100B+ annual revenue.
Anthropic $380B ~$3–5B est. Unprofitable Potential 2026 IPO; dependent on enterprise + API revenue scaling.
xAI (Grok) $50B+ Undisclosed Burning Leveraging Tesla / X infrastructure; no independent financial disclosure.
Mistral ~$6B Small Burning European positioning; open-weight models as strategic differentiator.
DeepSeek N/A (state-linked) N/A $6M training run Demonstrated that frontier training at 1/16th cost is possible. Changed the entire industry's assumptions.
✦ The OpenAI Math — And Why It's So High-Stakes
OpenAI's path to profitability requires revenues to grow from $13 billion in 2026 to over $100 billion by 2030 — an 8× increase in four years — while compute costs are simultaneously scaled up to $665 billion in total spending. For context: no software company in history has grown from $13B to $100B in four years from a standing start. Salesforce took twelve years to reach $26B. The thesis depends on agentic AI — AI that replaces entire workflows rather than augmenting them — being monetised at a scale that has not yet been demonstrated. The next 18 months of enterprise AI adoption data will be decisive.
Layer 5 Deep Dive

The Graveyard: 40% of AI Startups Gone in 24 Months

This is the most visible evidence that the AI bubble exists — and also the most misleading metric if taken in isolation. Most of the failed startups deserved to fail. The question is whether the survivors have real moats, or are just better-disguised wrappers.

💀
Failure Mode 1 — The Wrapper Problem
"GPT-4 + specialized prompt + nice UI" was a business model. Then OpenAI shipped the same features.
OpenAI's product cadence — GPT Store, Operator, Tasks, Canvas, Deep Research, Assistants API — directly cannibalized at least 200 funded startups in 2024 alone. Companies that had raised $10–30 million Series A rounds for "AI writing assistants," "AI research tools," and "AI coding tools" found those features shipping inside ChatGPT Plus at $20/month. Inference cost per million tokens dropped 80% from 2023 to 2025 — good for users, fatal for startups whose only margin existed between API costs and what customers would pay. The startups that built on OpenAI's API without proprietary data had no moat by definition.
200+ startups cannibalized by OpenAI product releases in 2024
95%
Failure Mode 2 — Enterprise AI Pilots Going Nowhere
95% of generative AI pilots are failing to scale, per MIT's 2025 study. BCG puts 74% of companies struggling to extract value.
The enterprise AI adoption story is not the story that press releases suggest. MIT's summer 2025 report found that 95% of generative AI pilots failed to generate meaningful business impact. BCG found that 74% of companies struggle to scale value from AI initiatives. IBM's CEO study found only 25% of AI initiatives deliver expected ROI, with just 16% scaled enterprise-wide. The pattern: companies ran experiments, demos looked compelling, scaling revealed integration costs, change management failures, and quality issues at production scale that weren't visible in the pilot. The $9 million annual cost of low-quality AI output (Harvard Business Review) is rarely included in AI project ROI calculations.
95% of pilots failing (MIT 2025) 25% of AI initiatives delivering expected ROI (IBM)
✓
The Survivors — What They Have That Others Don't
The AI startups that are succeeding share one trait: proprietary data in a vertical where foundation models won't go
The survivors are not randomly distributed — they cluster in regulated industries. Healthcare AI companies with proprietary diagnostic or clinical data. Legal AI companies with firm-specific document repositories. Finance AI companies with transaction data that cannot be replicated by a general model. Compliance and cybersecurity AI where liability concerns make foundation model companies reluctant to compete. The pattern: foundation models are extraordinarily capable but won't enter markets where a hallucination creates legal or financial liability. Vertical AI companies in those markets are insulated from commoditisation — for now.
Healthcare · Legal · Finance · Compliance
The Forward View

When the Math Has to Change — And Three Scenarios for What Happens Next

The AI economy's current equilibrium — infrastructure wins, hyperscalers spend, foundation models burn, startups fail — is not stable indefinitely. There are three plausible paths forward, with very different implications for every layer of the stack.

Scenario A — The Bull Case: Agentic AI Monetises
Enterprise AI agents replace workflows, not just tasks. Foundation model revenue grows 8× in four years. The capex is justified.
The bull case requires that AI agents — software that autonomously completes multi-step professional tasks without human oversight — generate revenue at a scale that justifies the infrastructure built to power them. OpenAI's Operator, Anthropic's computer use, and Google's Project Mariner are early demonstrations. If vertical AI agents replace entire job functions (a paralegal, a junior analyst, a customer service operation) and are priced accordingly — at $50,000–$500,000 per year per "agent employee" — the revenue math for foundation models begins to work. This is the scenario where current valuations prove conservative.
Scenario B — The Bear Case: Enterprise Adoption Plateaus
The 95% pilot failure rate becomes permanent. Hyperscalers slow capex. NVIDIA's order book stalls. The unwinding begins.
If enterprise AI delivers the productivity it promises in pilots but fails to scale — due to data quality, change management, or regulatory constraints — CFOs who approved AI budgets in 2024 begin demanding proof in 2026. Hyperscalers reduce capex growth. NVIDIA's $1 trillion order book faces cancellations. Foundation model companies cannot raise the next round at current valuations. CoreWeave's GPU-backed debt — rated A3 based on Microsoft and Meta creditworthiness — faces a repricing. 498 AI unicorns begin a multi-year valuation reset. This is the 2001 scenario: not a crash, but a long, painful unwinding.
Scenario C — The Most Likely Case: The Stack Separates
Infrastructure wins regardless. Foundation models consolidate to 2–3 survivors. The startup graveyard expands. Enterprise adoption is slower than projected but real.
The most historically accurate analogy is not the dot-com bubble — it is the railroad consolidation of the 1880s. Railway infrastructure was massively overbuilt. Most railway companies went bankrupt. But the railways themselves remained, carrying more freight each decade. In the AI version: NVIDIA, TSMC, and energy infrastructure win regardless of which foundation model survives. OpenAI likely survives — it has Microsoft as both customer and investor, and enough revenue traction to reach profitability if compute costs fall. Anthropic, xAI, and the others face a consolidation that reduces the current six major foundation model companies to two or three by 2028. Enterprise AI adoption continues but at the slower pace the actual ROI data suggests — generating real but below-projection revenue. The startups without proprietary data finish failing. The startups in regulated verticals build durable businesses. This is neither a catastrophe nor a vindication of current valuations — it is a messy, uneven transition that will look obvious in retrospect.
Interactive

Quiz: Where Does Your Money Sit in This Stack?

Your exposure to the AI spending cycle isn't just your stock portfolio — it's your employer, your 401(k), your unvested RSUs, and the industry your career sits in. Five questions to map your real position.

AI Spending Exposure Quiz
Which Part of the AI Money Stack Are You In?
Your result maps your career and portfolio exposure to each layer of the AI spending cycle — and what the three scenarios mean for you specifically.
1 / 5
Where does your income primarily come from?
Infrastructure or semiconductor company (NVIDIA, TSMC, cloud hardware, energy)
Hyperscaler or large cloud company (Microsoft, Amazon, Alphabet, Meta, Oracle)
Foundation model company or AI lab (OpenAI, Anthropic, xAI, or similar)
AI startup — application layer, SaaS, or AI-native product company
Non-AI enterprise using AI tools (finance, healthcare, law, consulting, etc.)
What does your equity / investment portfolio look like?
Heavily concentrated in AI stocks — NVDA, MSFT, or employer RSUs at an AI company
Mostly S&P 500 index fund — significant passive AI exposure I don't fully track
Actively diversified away from AI and tech concentration
Significant unvested equity at an AI startup or pre-IPO company
How dependent is your employer's business model on AI spending continuing to grow?
Very directly — our revenue follows AI capex (chips, cloud, infrastructure)
Indirectly — we sell to enterprises adopting AI, but we're not AI infrastructure
Mixed — AI is one growth driver among several, not the whole thesis
Low — we're in a regulated vertical where AI is a slow-moving tailwind, not a dependency
How does your role intersect with the 95% pilot failure rate for enterprise AI?
I sell AI products or services — the failure rate is my pipeline problem
I deploy AI inside my company — I'm in the middle of that 95% failure statistic
I use AI tools in my work — whether they show ROI affects my department's budget
My role doesn't depend on AI adoption metrics — I'm largely insulated
If hyperscalers cut AI capex by 30% in 2027, what would happen to your financial situation?
It would be severe — my employer's revenue or my stock portfolio depends on that spending
It would hurt — I'd see real impact on growth prospects or equity value
Minor impact — I'd feel it but it wouldn't be a crisis
Neutral — I'm not materially exposed to the hyperscaler capex cycle
–
Your AI Stack Position
Calculating…
Common Questions

FAQ

Parts of it are, and parts of it are not. NVIDIA's $120B GAAP net income ($117B non-GAAP) on genuine hardware demand is not a bubble — it is a real hardware cycle with real margins. The 40% of AI startups that shut down in 24 months is a classic bubble pattern. OpenAI's $730B valuation against $5B in annual losses is at minimum a valuation that requires extraordinary execution to justify. The bubble exists at the application and valuation layers — not the infrastructure layer, which is experiencing genuine, measurable demand that will persist regardless of whether any specific AI application succeeds.
The clearest winners are the hardware and infrastructure layers. NVIDIA ($120B GAAP net income, 71% full-year gross margin, $4.5T market cap). TSMC (the sole manufacturer of AI chips at leading-edge nodes; order backlog through 2027). SK Hynix (HBM memory monopoly on the highest-margin GPU configurations). Vistra and Constellation (nuclear power operators repriced by AI data center demand). The hyperscalers are generating real AI cloud revenue — Azure AI and AWS AI services are growing faster than their overall cloud businesses — but they are spending more than they're earning in AI specifically. The model companies and startup application layer are almost uniformly losing money.
Three causes dominate: (1) Foundation model commoditization — OpenAI's product cadence directly eliminated at least 200 funded startups in 2024 by shipping features that had been startup products; (2) No proprietary data moat — startups built API wrappers around foundation models with no unique data, making them obsolete with every model update; (3) Unsustainable unit economics — GPU compute costs of $1M+ per month, combined with declining willingness-to-pay as free alternatives proliferated, made positive unit economics impossible without a genuine data advantage. The survivors are concentrated in regulated verticals — healthcare, legal, finance, compliance — where foundation model companies won't compete due to liability concerns.
At current revenue of approximately $13 billion, the $730 billion valuation implies a ~56× revenue multiple — extraordinary even by high-growth software standards. Justifying it requires OpenAI's revenue to grow from $13 billion to over $100 billion by 2030, which would be the fastest revenue growth in software history from that base. The bull case depends on agentic AI — AI agents that replace entire job functions and are priced accordingly — achieving genuine enterprise scale within four years. The bear case is that enterprise AI delivers productivity gains but not the workflow displacement required for that revenue scale, and OpenAI's valuation corrects substantially. OpenAI is not worthless — it has the most capable AI products, the largest distribution, and the most recognisable brand in the market. But at $730 billion, the margin for execution error is extremely small.
NVIDIA is structurally exposed to hyperscaler capex — approximately 70–75% of its data center revenue flows from Microsoft, Amazon, Alphabet, Meta, and Oracle. If hyperscalers slow capex growth materially, NVIDIA's revenue growth slows or contracts. The January 2026 DeepSeek shock — which briefly wiped $600 billion from NVIDIA's market cap in a single day — showed how sensitive the stock is to any narrative that reduces compute intensity. The underlying demand driver is more durable than the headlines: AI inference token generation has grown tenfold in a single year, and agentic AI is compute-intensive. The risk is not demand disappearing — it is demand growing more slowly than the current valuation requires.
Disclaimer: This article is for informational purposes only and does not constitute financial, investment, or career advice. All data points are sourced from publicly available earnings filings, research, and news reports as cited. Company financial figures reflect reported or estimated results as of Q2 2026 and are subject to revision. This is not a recommendation to buy, sell, or hold any security. Forward-looking scenarios are editorial analysis, not forecasts.
References

Sources & References

1
NVIDIA SEC Filing (Feb 2026). NVIDIA Q4 FY2026 Earnings; Al Jazeera (May 21, 2026). Nvidia posts record profit of $58.3bn. aljazeera.com ↗; Fortune (Feb 2026). Nvidia smashes Q4 2026 with $68 billion in revenue. fortune.com ↗
2
No One's Happy (May 2026). The AI Bubble. CoreWeave $8.5B term loan rated A3 by Moody's. nooneshappy.com ↗; Consumer Reports / IEA (2026). AI Data Center Power Demand Projections.
3
Futuriom (Apr 2026). Hyperscaler AI Spending Doubts Rising. Oracle capex 66% of revenues. futuriom.com ↗; Medium / Aftab (Apr 2026). The $2 Trillion AI Bubble. Morgan Stanley capex-to-sales ratio analysis.
4
UnboxFuture (May 2026). The 2026 AI Bubble Burst: When Subsidies End. OpenAI-Oracle financing structure; Columbia Business School Prof. Van Nieuwerburgh quote. unboxfuture.com ↗; MEXC (Feb 2026). OpenAI Projected to Post Over $14 Billion in Losses in 2026.
5
IBM Think (Jun 2026). How to Maximize AI ROI in 2026 — MIT 95% pilot failure rate; IBM CEO study. ibm.com ↗; IdeaProof (2026). 319+ AI Startups That Failed. CB Insights / Gartner 40% failure rate data. ideaproof.io ↗; Medium / AI Empire Media (Mar 2026). The Real Reason AI Startups Are Failing in 2026.
6
Kannan K R / Medium (Mar 2026). The AI Bubble: Hype, Capital, and the Trillion-Dollar Question. JP Morgan Asset Management data: AI stocks 75% of S&P 500 returns since ChatGPT. medium.com ↗
7
Kiplinger (May 2026). Nvidia Earnings: Updates and Commentary. Trillion-dollar order book; Blackwell ramp. kiplinger.com ↗
8
MarketWise (Feb 2026). Hyperscaler AI Investment to Surge 71% in 2026. Alphabet 100-year bond; Oracle $25B stock issuance. marketwise.com ↗
Written by
HenryPulse Research & Editorial Team
HenryPulse produces data-driven financial content for high-income professionals in tech, finance, and strategy. Our research combines public market data, earnings disclosures, and third-party industry reports.
Not financial advice. This article is for informational and educational purposes only. It does not constitute financial, tax, investment, or legal advice. Always consult a qualified professional before making financial decisions. Full disclaimer →