The bubble argument against generative AI’s viability arrives on schedule every few months, and it will keep arriving. It deserves a definitive answer rather than another round of reassurance. The answer is that the valuations rest on a real and growing revenue substrate, that the substrate predates the generative AI wave by two decades, and that the failures now being catalogued are deployment failures at the application layer rather than evidence that the underlying economics are fictional. An industry does not dissolve because one accounting assumption about chip life turns out to be too generous.
The right conclusion is not that there is no excess in AI. There is excess. The right conclusion is that excess is forming on top of a real economic base, not in place of one. That makes this a repricing problem a deployment problem, and an accounting-quality problem. It does not make the entire AI economy fictional. When tracing the actual history, because the history is the argument, the technology that markets are now pricing did not appear with ChatGPT. It has been compounding inside the global economy since the early 2000s, mostly invisibly, mostly profitably. Anyone who says that AI is about to burst is either committed to a bit, committed to a position they can no longer separate themselves from, or don’t understand the history of AI or bubbles. The negativity many are focusing on is a manifestation of animosity and almost wishful bubble thinking, but the bubble itself is not real. Not in the traditional sense and not in the revenue sense. Is there froth in the system? A lot. Of course, the way there is with any new technology, and there’s never been a technology like artificial intelligence, but that doesn’t mean that the industry is failing.
If we look back at the history of generative AI and how it evolved out of machine learning, an evolution which has transformed virtually every industry it has touched (for good or for ill) the true picture emerges
Background: Twenty Years of Machine Learning That Paid for Itself
Machine learning became a foundational commercial technology through advertising, and it did so long before anyone in a boardroom used the word “AI.” The earliest deployments automated pay-per-click bid management. As data pipelines and model quality improved, that narrow task expanded into programmatic real-time bidding, where platforms like The Trade Desk, Google Ads, and MediaMath place and price ads in milliseconds across billions of impressions. The same statistical machinery moved into recommendation engines. Netflix and Amazon built recommendation models on known customer attributes and habits, then used those models to suggest content and products to new customers. Search engines and social networks that live on ad revenue ran the same playbook.
This layer worked. It generated measurable return for twenty years, it survived the 2008 crash and the dot-com aftermath, and it scaled into the connective tissue of digital commerce.
The crucial point for any bubble discussion is that the recommendation and targeting substrate is mature, proven, and still expanding. It is the floor beneath everything built since 2022. When critics describe the entire AI economy as speculative, they are describing froth on top of a deep base while pretending the base is not there.
Adoption then spread industry by industry. Dynamic pricing in airlines and hotels, demand forecasting in retail, fraud detection in payments, churn prediction across subscription businesses. None of this depended on the transformer architecture. All of it established that statistical learning over business data produces durable economic value when the data is good and the problem is well defined. Those two conditions, good data and a well-defined problem, will return as the decisive variables in the generative era.
The GPT Inflection and the Generative Explosion
The public release of large language models changed the interface, the addressable surface, and the capital flows. Generative systems extended machine learning from ranking and prediction into language, code, and synthesis. The adoption curve that followed has been faster than any prior enterprise technology.
The Canadian data makes the speed concrete. A BDC survey released today of 1,500 business owners found that generative AI, which was not yet available to the general public in the 2021 version of the survey, is already used by three out of ten SMEs. A technology that did not exist in consumer form four years ago now sits inside roughly a third of Canadian small and medium businesses. The same survey found that 78% of SMEs using generative AI are satisfied with the return on that investment.
That satisfaction figure is the part the bubble thesis cannot easily absorb. Speculative manias do not usually produce majorities of paying users who report that the spending was worth it. The return is showing up in operating results: the survey attributes improved operational efficiency to 43% of surveyed businesses and improved client experience to 34%, and its econometric work finds that AI use is associated with a productivity level roughly 24% higher once company characteristics are controlled for.
The Capital Signals: IPO and Equity Raise on the Same Day
Two events on June 1, 2026 frame the financial question. They should be read together.
Anthropic confidentially submitted a draft S-1 to the SEC on June 1, 2026 for a proposed initial public offering. The filing followed a $65 billion Series H round closed four days earlier at a $965 billion post-money valuation, surpassing OpenAI’s $852 billion valuation disclosed in March. The valuation trajectory is steep on its own terms: the company was valued at $380 billion in February 2026, and by late May that number had more than doubled to $965 billion.
On the same day, Alphabet announced its first equity raise in two decades. Alphabet announced equity offerings totaling $80 billion in expected aggregate amount, as part of its plan to fund investments in AI compute infrastructure to meet customer demand, including an agreement by Berkshire Hathaway to invest $10 billion in a private placement. The stated reason matters: the company said it is experiencing strong demand for its AI solutions from enterprises and consumers at levels exceeding its available supply. Alphabet’s 2026 capital expenditure guidance sits at $180 billion to $190 billion, roughly double 2025’s $91.4 billion, and Google Cloud reported a 63% revenue jump with a $460 billion contract backlog.
A demand-exceeds-supply constraint is the structural opposite of a bubble. A bubble is too much capital chasing too little real use. Alphabet is describing customers it cannot currently serve and a backlog measured in hundreds of billions of dollars. The Berkshire Hathaway participation belongs in the same reading. The most valuation-disciplined institutional investor in the world anchoring a $10 billion placement is not the behaviour associated with the top of a mania.
The Anthropic valuation deserves scrutiny rather than applause, and the scrutiny strengthens the case. The figures only hold if revenue growth and a credible path to profitability justify the price. On that, Anthropic projects breaking even by 2028, two years ahead of OpenAI’s 2030 profitability target. Reported revenue run-rate figures vary widely across outlets and should be treated with caution until the S-1 detail becomes public. What is verifiable is the customer trajectory: business customers exceeded 300,000, up from fewer than 1,000 two years prior, with large accounts growing more than sevenfold in a year. That is enterprise revenue formation, not retail speculation.
The Failures Are Real, and They Are the Wrong Failures for the Bubble Case
A serious argument has to hold its own counter-evidence in view. The deployment failures are genuine, they are instructive, and they have been a throughline of my published work on agentic systems. They simply do not support the conclusion the bubble case needs.
Starbucks is the clearest retail example. The company terminated its AI-powered inventory counting system across all North American stores in May 2026, nine months after deploying it across more than 11,000 company-operated locations. The system had been announced with a claim of 99% accuracy. On the floor, employees described it frequently miscounting and mislabeling items, confusing similar milk types, missing items during scan sessions, and in one case failing to recognize a syrup bottle in a promotional video Starbucks itself had uploaded to showcase the tool. The diagnostic detail is the one that matters: a system that requires workers to verify every output delivers no net efficiency gain, because it doubles the task.
Two things about the Starbucks reversal are easy to miss. First, the company did not retreat from AI. It kept Green Dot Assist, the barista support chatbot, in active rollout. The retired system was the one operating in the physical world with no human review between output and consequence. The retained system is the one where an imperfect AI suggestion can be reviewed and overridden before it causes a downstream problem. The line between those two cases is the line between deployments that work and deployments that fail. It is a line about the harness around the model, not about the model.
Salesforce shows the same pattern at the enterprise layer. Bloomberg reporting raised pointed questions about the gap between Agentforce marketing and Agentforce delivery, including an advertisement depicting a hospital system where the agent books appointments and arranges refills, capabilities that Bloomberg reported are still unavailable, with patients today greeted by keypad menus and routed to human schedulers, and the chatbot still being tested and not visible to most web visitors. Salesforce’s own engineering writing describes the failure mode precisely. An agent can return a plausible, well-formed response that is completely wrong for the situation, with no error thrown, no alert fired, and nothing in the logs to indicate a problem; the failure mode is semantic, not technical. Salesforce locates the cause where the evidence points: the most consequential factor that determines whether an agent succeeds is not the model powering it but the architecture built around it, the data it can see, the permissions it operates under, and the systems it can reach.
Amazon’s internal AI leaderboard offers the same lesson from inside the company doing the spending. The dashboard, known internally as KiroRank, ranked developers by how much they used Amazon’s Kiro AI coding platform, measured in raw token consumption. Employees did what any incentive structure invites: they gamed it. Some ran scripts and spun up pointless agents to inflate their usage scores, with one employee saying they cheated after management told them they were not using AI enough. The volume metric rewarded volume, and burning tokens carries real cloud cost, so the program generated expense without generating work. Amazon deprecated the leaderboard and replaced it with a metric it calls “normalized deployments,” which tracks AI-generated code that actually ships and functions. The senior executive overseeing the change, Dave Treadwell, reportedly told staff to stop using AI for the sake of using AI. This is Goodhart’s Law operating exactly as predicted: once a measure becomes a target, it stops being a good measure. What it is not is evidence that the underlying tool fails. Amazon’s correction was to stop measuring activity and start measuring useful output, which is a refinement of how to capture AI value, not a retreat from the claim that the value exists. The same company killing the vanity metric is targeting more than 80% weekly developer AI adoption and spending roughly $200 billion in 2026, most of it on AI infrastructure.
(Each company mentioned here also has significant holdings in Anthropic. Salesforce owns about 1% of Anthropic, implying a stake worth roughly $9.65B at the reported $965B valuation, though Reuters recently referred to Salesforce’s stake as $5B before that latest mark. Amazon’s Anthropic holdings were reported at $74.2B, implying about 7.7% of Anthropic at the $965B valuation. Google/Alphabet is reported to be allowed to own up to 15% of Anthropic, which would imply up to about $144.75B at a $965B valuation, though its exact current percentage is not clearly disclosed.)
Microsoft has also pulled back from Anthropic use because of cost, while yesterday in introducing seven of its own new AI models, including an agentic offering based on OpenClaw. This is either the pilot-to-production gap or a competitive decision. Neither indicates a bubble. A company like Microsoft rarely doubles down on an enormous technology that has not proven its value, and the company is notoriously late to enter the game. It sits on the sidelines and wait to see what happens first. By jumping in with seven new models yesterday it demonstrates it’s committed.
Production is the central deployment failure site for AI in the enterprise. The model demonstrates capability in a controlled pilot. Production introduces messy data, broad permissions, real consequences, and no semantic monitoring, and the same model produces confident wrong answers at scale. The BDC data confirms the diagnosis from the other direction. Satisfaction with the return on AI investment is much higher among SMEs with fully integrated data at 94%, than among those with partially integrated or unintegrated data at 63%, and almost half of SMEs with digital data have data that is not integrated at all or only very partially integrated. The technology delivers when the data substrate and the surrounding architecture are sound. It disappoints when they are not.
These failures kill a specific overclaim, that you can drop a frontier model into an unprepared organization and harvest autonomous productivity. They say nothing against the proposition that AI generates large, real value where it is deployed against good data and a defined problem. A bubble bursts when the underlying value was never there. What Starbucks and Salesforce demonstrate is that the value is conditional, which is a very different and far more survivable claim.
Uber’s recent decision to limit token spending to $1500 per employee per month per tool has provoked questions like “why is Uber still coding?” (tech companies are always coding) and statements like “look it’s … you guessed it … a bubble.” I would posit that a $1500 a month cap on spending PER TOOL on a relatively new technology is the opposite of proof that companies are retrenching. What they are doing is limiting their expenditures on something that very quickly has gotten out of control for others.
China: Mass Adoption at Scale, With Real Caveats
China is the clearest national case of treating AI as economic infrastructure rather than a speculative asset class. The 15th Five-Year Plan, approved on 12 March 2026, mentions AI 52 times, up from 11 in its predecessor, gives computing power its own dedicated chapter, and frames an “AI+” initiative targeting integration across 90% of the economy by 2030. The plan authorizes a procurement and subsidy architecture, including government procurement of compute services and greater procurement of indigenously innovated products.
A state planning a near-total diffusion of AI across its productive economy is making the same bet the capital markets are making, through a different mechanism. When both the most market-driven actors and the most state-driven actor are committing at scale, the simplest explanation is that the underlying technology produces value, not that two opposite systems are simultaneously deluded.
The caveats are real and belong in any honest account. Chinese adoption runs on a deliberately different cost structure: rather than competing frontier-to-frontier for the most capable closed models, Chinese developers have pursued an open-source, low-token-cost approach. That strategy is partly a response to constraint rather than pure preference. The plan is also a self-reliance document built under export pressure: it codifies a domestic equipment rule for chip manufacturing and a ban on foreign AI accelerators in state-funded data centers. The economics are further flattered by direct subsidy: the Ministry of Finance has proposed a 20% discount for domestic products in government procurement, and energy subsidies have reportedly cut power costs by up to half for some data centers using Chinese-made processors. Subsidized power and protected procurement mean the reported efficiency of Chinese deployment is not a clean market signal. The adoption is genuine and the scale is real. The unit economics are partly policy-driven rather than purely competitive, and the 90% target is an aspiration in a planning document rather than an achieved figure.
The Valuations Track Fundamentals, Not Sentiment
The bubble thesis claims valuations have detached from fundamentals. The evidence points the other way. Valuations are tracking a real and growing revenue base, demand that currently exceeds supply, and an adoption curve faster than any comparable enterprise technology, while the visible failures are concentrated at the deployment layer rather than the demand layer.
The strongest technical version of the bubble argument does not rest on sentiment at all. It rests on depreciation, and it is worth taking on directly because it is the most serious objection.
The Depreciation Problem, Taken Seriously
Michael Burry has made the most rigorous bear case, and it should be stated at full strength. He argues that hyperscalers flatter earnings by depreciating Nvidia-based hardware over five or six years even though Nvidia’s fast chip cycle means the real economic life is closer to two or three years, producing an estimated $176 billion of understated depreciation and overstated profits across the industry between 2026 and 2028. The mechanism is correct as far as it goes: extending an asset’s assumed life lowers annual depreciation and raises reported profit, and Nvidia’s release cadence has accelerated to roughly annual, with each generation delivering 40 to 50% better performance per dollar, which does erode the economics of keeping old chips at the training frontier.
The argument has two real weaknesses, and the cited evidence supplies both.
First, it conflates the training frontier with the economic life of the hardware. The obsolescence cycle for frontier training, where chips age out in 18 to 36 months, is distinct from the economic utility cycle, in which a chip drops from frontier training to cheaper inference and continues generating revenue for years. A GPU that is no longer competitive for training the newest model is still productive serving inference, and inference is where the volume demand is heading. CoreWeave has cited real-world data showing that older Nvidia A100 chips retained 95% of their original price in expired contracts, which is hard to reconcile with a two-year economic death. Oracle, Meta, and Microsoft have to defend these schedules to auditors with utilization and failure data, and they keep passing.
Second, and decisively for the bubble question, depreciation is a non-cash accounting entry. A company’s free cash flow is identical whether it depreciates over three years or six, because the money left the building when the GPUs were purchased. The depreciation schedule changes reported accounting profit. It does not change the cash that has already been spent or the revenue the hardware earns. The honest conclusion is the narrow one: if Burry is right, reported earnings at some hyperscalers are overstated for a few years and will be revised as schedules tighten. They will need to spend more more frequently on newer chips. That is great news for NVIDIA and an earnings-quality issue and a reason for analytical care on specific names. It is not a mechanism that vaporizes demand, erases the 300,000 enterprise customers, or unwinds a $460 billion cloud backlog.
Even Nvidia engaged the charge directly rather than ignoring it. The company issued a detailed memo to Wall Street analysts disputing the claim and stating that its business is economically sound and its financial reporting transparent. Whether one fully accepts that rebuttal, the point stands: the depreciation debate is a debate about the slope of reported profit, not about whether the revenue and the demand are real.
An Industry Does Not Dissolve Over an Accounting Assumption
Put the pieces in one frame. A machine learning substrate that has paid for itself for twenty years through advertising and recommendation. A generative layer adopted faster than any prior enterprise technology, with a majority of business users reporting positive return. Demand that the largest infrastructure provider in the world says exceeds its supply, funded by its first equity raise in two decades and anchored by the most disciplined value investor alive. A near-trillion-dollar IPO filing backed by enterprise customer formation rather than retail froth. A state planning to diffuse the technology across nearly its entire economy. And against all of that, a set of failures concentrated precisely where theory predicts they would be, at the deployment layer, in organizations with unintegrated data and unmonitored agents, alongside an accounting dispute about the timing of non-cash charges.
The deployment failures are the most useful information in the whole picture, because they tell builders and buyers exactly where the work is: data integration, architecture, governance, and human review between output and consequence. They are the conditions of value, not evidence of its absence.
A bubble is a structure with nothing underneath it. This structure has twenty years of compounding machine learning underneath it, a revenue base that is growing, and demand that currently outruns supply. The chip depreciation question is worth getting right, and it changes some reported profit figures for a few years. An entire industry does not dissolve because depreciation schedules were set longer than they should have been. The fundamentals are doing the work the bubble narrative attributes to sentiment. What the industry is missing is regulation that enables people to understand what the true limitations of the technology are, and the presidential EO yesterday did little to change that.
Jen Evans is Principal of Pattern Pulse AI and co-founder of Tech Reset Canada. This piece builds on her ongoing B2BNN research into LLM failure modes, agentic systems, and the pilot-to-production gap.