On 29 April 2026, four of the five most valuable companies in the world reported earnings inside a 19-minute window. The numbers were the largest in tech history.
For the quarter, Google Cloud grew 63% to $20 billion. AWS grew 28% to $37.6 billion, the fastest in 15 quarters. Azure grew 40%.
Combined, the three big clouds reported $1.45 trillion ($627 billion at Microsoft, over $460 billion at Google, $364 billion at AWS, before counting Anthropic's $100 billion-plus April commitment) in audited contractual backlog, or "remaining performance obligations" (RPO), which signals future revenue from AI-driven cloud contracts, verified through audits under standards like ASC 606.
2026 capex guidance across the four hyperscalers and Meta now sits at roughly $725 billion. That is more than 2% of US GDP, deployed by five companies, in twelve months. That number could hit $1T in 2027.
This is not what a bubble looks like. The harder question is why so many smart people thought it was (and still is).
Six structural reasons this isn't a bubble
The bear case in 2024/2025 was reasonable. Sequoia's David Cahn asked the "$600 billion question": what revenue justifies the GPU spend? Goldman Sachs ran a house view asking "what trillion-dollar problem is AI solving"? Bloomberg published a graphic-led investigation into the closed loop of hyperscaler-to-frontier-lab funding (below).
I've never been sure the argument carried much weight, largely subscribing to Azeem's Azhar view of the world. The problem was the AI bubble debate took-over conversations for a large period of time, when in reality the impacts of any fall out, I believe, wouldn't have a material impact on the average Joe.
A few areas that have hardened my POV that we're not in a bubble:
**The OpenAI-concentration concern has been resolved by diversification.**Not by OpenAI accelerating, but by the revenue base broadening, and other labs catching up in capability and capturing new and existing user demand. Anthropic went from $1 billion ARR at the start of 2025 to $30 billion in April 2026 (gross figure; OpenAI's CRO disputes the methodology and puts the net-comparable figure closer to $22 billion). Either way, 8 of the Fortune 10 are paying customers, and 1,000+ companies spend over $1 million annually with Anthropic alone. Google Gemini Enterprise paid MAUs grew 40% QoQ. Total enterprise generative AI spend tripled from $11.5 billion in 2024 to $37 billion in 2025 (Menlo Ventures).
**The companies aren't debt-levered yet. Not borrowed money funding the build-out.**Operating cash flow at Amazon is healthy at $148 billion trailing twelve months, up 30%, and capex is being funded from current cash. Free cash flow has compressed because all of it is being redirected into infrastructure (Amazon TTM FCF $25.9 billion to $1.2 billion). A capital structure observation, not a solvency one.
**The dollars flow downstream into the wider economy. Not a closed loop.**Bloom Energy is up roughly 1,400% on the year on hyperscaler power demand. SanDisk is up 30x in two years on memory supply pressure. Energy producers, memory makers, networking gear, data-centre construction firms: a trillion dollars a year is leaving hyperscaler balance sheets and entering the broader economy. NVIDIA's top four customers today (the hyperscalers) generated $451 billion in operating cash flow in 2024 from real paying customers. Different category.
Customers are paying and locking in 12 to 18 months early. $1.45 trillion in signed multi-year backlog.
**Demand isn't even at full intensity yet.**Alongside the OpenAI diversification point, this is the most important point (referred to here six months back). Jensen Huang on BG2 Pod articulated the mechanism: AI demand is two stacked exponentials, more users multiplied by more compute per user, with token generation doubling every few months. Critically, on the adoption curve, 1) Agents aren't running 24/7 at scale yet, 2) most enterprise users haven't moved to reasoning models or leveraging multi-modal capabilities, which consume an order of magnitude more compute per query than chat. Each shift is a multiplier still ahead of the curve, the Q1 numbers are the lowest floor we'll see, IMO.
**Supply is the constraint, not demand.**Three CEOs said it directly on the Q1 calls. Sundar Pichai: "we are compute constrained in the near term, and our cloud revenue would have been higher if we were able to meet that demand". Amy Hood described Azure as capacity-constrained through 2026. OpenAI CFO Sarah Friar called it "a vertical wall of demand right now". In every previous tech bubble, supply chased phantom demand. In Q1 2026, supply is still catching up to signed demand.
What happened to the bubble narrative
The "is this a bubble?" question doesn't really come up in industry or customer conversations any more. It just isn't there, not because anyone declared the debate over, but because the market has matured. Customers have moved past the question of whether AI is a fad and into the question of how to make it work in their specific business. The framing shifted from existential ("should we?") to operational ("how do we, without wasting capital?"). The data matured, and the conversation matured with it.
That shift is doing work the bear case can't. AWS Bedrock processed more tokens in Q1 2026 than in all prior years combined. 125,000 customers, +170% QoQ Bedrock spend, almost 80% of the Fortune 100 (Jassy verbatim). Google Cloud's operating margin expanded from 17.8% to 32.9% in a single year while capex accelerated. The TPU bottleneck is its own directional signal. Enterprise compute is flowing to Google because Nvidia is sold out, and Anthropic's $100 billion-plus Broadcom-Google TPU deal is enterprise demand literally building around the chip shortage.
The risks that did survive Q1
The challenge is that two real risks remain on the board pack, and they are not the bubble risk the narrative gravitated towards.
**1. OpenAI-specific cash burn is now a stock-picking risk, not a sector risk.**Annualised revenue grew from about $20 billion in November 2025 to $25 billion in February 2026, healthy on absolute terms but missing an internal stretch target. Projected 2026 losses sit at $14 billion against $17 billion of cash burn (per The Information), with no break-even until 2029. The company is structurally dependent on continued external funding closing on time. It just closed a $122 billion round in March at an $852 billion valuation.
The challenge is who owns that risk. OpenAI has no ticker and no quarterly disclosure. But on 27 April a WSJ leak about missed targets moved Oracle by up to 5.5% in a day, NVIDIA by 3%, AMD by 3 to 4%, Broadcom by 4%, SoftBank by 10%, and CoreWeave by 5.4%. Stocks recovered within 48 hours, but the move proves the exposure. Any board or institutional fund holding NVDA, ORCL, AVGO, SoftBank, CoreWeave, or Microsoft has indirect exposure to OpenAI's strategic position, without OpenAI's disclosure obligations. NZ Super and Kiwisaver portfolios carry this exposure today.
2. The real AI risk is the cost of building it, the capital structure. In plain English, the big AI companies now need to spend huge money on data centres, chips, energy, infrastructure, and financing. Chamath Palihapitiya on All-In named it: "the pendulum is swinging violently from asset-light to asset-heavy". Bank of America has hyperscaler debt issuance hitting $175 billion in 2026, against a $28 billion annual average over the prior five years (roughly six times higher). Microsoft's $190 billion 2026 capex guidance includes about $25 billion explicitly attributed to higher component pricing. The risk isn't demand vanishing, it's demand softening for two quarters while debt service compounds and above-market Power Purchasing Agreements lock in, and debt costs compound. Long-term energy contracts stay locked in., while nargins get squeezed. This is now a question of profitability and cash flow, not whether the AI market is real.
Apple is the dog that didn't bark in this debate. Apple's Q1 capex spend was $1.9 billion against Amazon's $43 billion, Google's $35.7 billion, and Meta's $19.84 billion. R&D up 33% to $11.4 billion. New $100 billion buyback authorisation. Stock up 3.24% on the print, while Microsoft fell 4.4% and Meta fell 8.55%. Apple chose to license Gemini from Google for $1 billion a year rather than build a frontier system. The market priced both strategies positively the same week. Not a binary "all-in or lose" choice. There are now two valid paths: be a heavy-capex AI provider and absorb the debt-load risk, or be an asset-light AI customer and license a frontier provider.
What needs to land next
For CIOs, map every model-tier dependency to a primary and a fallback vendor across OpenAI, Anthropic, and Google Gemini. Renegotiate compute contracts on a 12 to 18-month lead window. Have a crisp answer to one question before the next board meeting: what happens if OpenAI restructures another partnership?
The execution gap is where the value lives now because Q1 settled the demand question. It didn't settle the harder one, which is whether companies can extract value from AI investments at the rate they're committing capital to them. The NBER data, the NANDA failure rates, the 42% of companies that abandoned most AI initiatives in 2025: none of that got refuted in Q1, because none of it was a bubble argument. It's an execution argument, and it's where the next 12 months of strategic discussion takes place.
The bubble question stopped showing up in customer conversations because customers have already started asking the next one. The boards still on the old question are looking backwards. The boards on the new one are doing the work that determines whether the bet is worth taking.
Passionate about all things AI, emerging tech and start-ups, Mike is the Founder of The AI Corner.
Subscribe to The AI Corner
