Are We In a Bubble

The Financial Times noted that the recent rally has been “driven by the smallest number of stocks on record.” In Korea, Samsung and SK Hynix were up 15.6% and 11.4% respectively on the announcement Apple may diversify its supply chain. The rally tripped circuit breakers. Jamie Dimon of JP Morgan says he loves vibe coding. He had Claude make him a mini dashboard to help him track rates. Everywhere you turn, someone is being creative and using the tools in a cool way, in many use cases.

However, the ebullience has the same feeling as past bubbles like the dot.com era. I remember having a dinner in 1999 and someone on the table telling me how the Internet could do everything, and he put all his savings in Yahoo. He said no one will visit a physical store in a few years. 25 years later, I still can’t get my wife to use Ocado, or online grocery delivery.

I’m part of a WhatsApp chat dedicated to AI and market themes that regularly debates ideas. When the Iran War was raging there were dramatic opinions and argumentative orations. Everyone had a distinct opinion. Recently, one member posted a chart from social media showing that 50% of the hyperscale capacity is essentially coming from OpenAI and Anthropic commenting “this won’t end well”. I responded by saying “based on Anthropic’s revenue and spend, and rapidly growing enterprise growth, it feels like those revenues are secure. OpenAI’s strategy is in flux so not sure about them. Meanwhile it seems AI use cases are rapidly growing. People who were sceptical now love vibe coding and are feeling an epiphany; not sure how much revenue that translates into, but the buzz is real." Given the group, I expected a vociferous counter-response about circular finance. Instead, it was crickets meaning I was either the first to win a debate in this WhatsApp chat, or more likely people just don’t know.

The way people are investing makes things feel like a bubble, but we can’t refute that AI is a revolutionary technology that’s growing at hyper speeds and is very useful.

So, I will summarize both sides of the debate and let readers decide.

The Future Is Here

Vinod Khosla, a venture capital investor, gave his data about the exponential growth of API calls he tracks. Andreesen Horowitz notes that the AI spend is not debt funded but coming from healthy balance sheets, suggesting past bubbles were usually created from excessive debt creation. They noted that the social media, mobile, cloud and SaaS cycle created $25 trillion of wealth. AI is destined to be more and only $7 trillion of market cap was added last year on $1.3 trillion of spending. OpenAI had a 300% growth rate and grew to $20 billion of revenue. Anthropic is now saying they will hit a run-rate of $30 billion and that use is up 80x. Most tech visionaries say that the AI will diffuse faster as it’s built on the rails of cloud and mobile. It will bend the productivity curve and bring the local surplus and economic growth throughout the world. Amazon, Meta, X.ai, Anthropic and ChatGPT are run by some of the most brilliant people in the world and they seem to believe that some sort of Artificial General Intelligence (AGI) is possible.

Andreesen noted the sharp pull of enterprise and unprecedented consumer adoption. The 1,000 enterprise clients spending over $1 million annually at Anthropic aren’t just buying tokens; they are integrating AI into daily workflows, retraining staff and building dependencies. This will extend revenue duration as enterprises lock-in.

Over the past two years, token costs have fallen 99%. Rather than killing margins, this is Jevons Paradox operating in real-time. AI unlocks applications that weren’t economically viable at $10 per million tokens. As marginal costs continue to lower, that is often a precursor to explosive adoption, which we are seeing real-time.

In the This Day in AI podcast, a programming-oriented podcast run by two Australian tech founders, for the past year they said Claude Code had not improved much, although it was packaged better. But in the last three weeks they noticed a step-change in Opus 4.6 and now GPT 5.5. AI coding tools are now accelerating AI development itself. Claude Code helps engineers build the next Claude faster. This recursive element means the capability curve may steepen even without fundamental architecture breakthroughs.

There were concerns that chips depreciate quickly, but many GPUs can run 5 to 7 years. Once the AI demand curve ends, these older GPUs could lose economic value quickly, but there is such demand that older GPUs are still being rented out.

At Davos, Satya Nadella noted that the adoption of AI must be more widespread than just the tech sector. Recent data-points show major progress in healthcare, finance, manufacturing, legal, consulting and accounting services. TCI, a hedge fund investor, sold shares in Microsoft allegedly on the perspective that their revenues from software may be compromised by AI.

There was also concern that ChatGPT and Anthropic could run out of money, but given they are raising billions at close to trillion-dollar valuations, they could have a few years of runway.

The above suggests that token demand and use will be insatiable. We are early in the growth curve, and we’ll have to build data centres in outer space and continue to produce more chips to satisfy all our token demand.

This is Going to End in Tears

Anthropic’s Dario Amodei says that in five years, AI will replace 50% of entry-level white-collar jobs pushing unemployment to 10%-20%. If that’s true, then voters will vote for politicians who limit AI.

Ken Griffin of Citadel notes that AI executives must push the narrative of artificial general intelligence (AGI) because otherwise they can’t raise the billions needed to buy GPUs and train models. One social media post from a prominent VC investor noted that no one is doing due diligence these days. Companies have no idea if the AI labs are copying their products. They are also tokenmaxxing, or where AI start-ups artificially inflate their token consumption to make it look like their usage metrics look larger than what genuine demand warrants. CNBC, TechCrunch and the NY Times also noted the phenomena of tokenmaxxing by start-ups, a real-world application of Goodhart’s Law, when a measure becomes a target, it ceases to be a good measure.

A Stanford report noted that 53% of ChatGPT’s users don’t pay anything. As AI labs force payment or seek monetization, consumers may defect to Chinese labs.

The This Day in AI podcast noted that agents haven’t advanced much from last year, the alleged ‘Year of the Agents’. Some smart people believe that new architectures are needed to advance the field and you can’t just vibe code your way to nirvana. These start-ups talk about world models, particularly to advance robotics.

A short seller report from a brokerage noted ugly cash generation, circular financing deals, rising component costs and big numbers attached to supply deals that are not what they seem. The Information said that in the Broadcom/OpenAI deal, nothing is as it seems. Softbank downsized a $10 billion margin loan on their OpenAI stake as lenders were only comfortable on $6 billion signalling a lack of confidence in recent funding round valuations. OpenAI is reportedly on track to spend $50 billion on compute in 2026 while generating perhaps $24 billion in revenues. The Stargate and Microsoft deals are the only thing making this math work, carrying concentration and counterparty risk.

Anthropic’s $30 billion run-rate is overwhelmingly driven by Claude Code. There is a risk agentic coding commoditizes faster than expected, especially as open-source alternatives from Qwen and DeepSeek improve. This also leads to the “10x developer productivity” paradox. If AI makes every software engineer 10x more productive, the rational enterprise response is to hire 90% fewer engineers. Tech companies have indeed announced significant engineer layoffs. This deflationary force hits the cohort that is AI’s most enthusiastic adopter and largest revenue source.

Chinese models are not far behind and are just as good as the Silicon Valley models and are selling tokens at 5% to 10% of the labs. Having tested DeepSeek’s new model on some of my workflows, I concur they are very good models. They are also efficient. Anthropic will reason out tokens such that I now put at the end of each prompt: “Limit your output to X words.” With DeepSeek, it gets down to what you want to know. I haven’t tried their coding architecture, but allegedly it’s competitive for module workflows.

The most compelling argument is that multiple U.S. states and European jurisdictions are actively restricting new data centre construction on water and power grounds. If electricity and water consumers stop data centre buildouts, the AI labs will have to be much more efficient on tokens or raise prices. Agentic flows will slow because people won’t want to pay the costs. As Elon Musk noted, the AI labs will see that software will soon learn about hardware by the end of the year as they hit physical barriers on data centre expansion. Chinese labs are also underpinned by a country that has invested significantly in renewables, grid and energy leading to a potential advantage down the road. The Chinese AI companies listed on the A-share market have had similar share price rises as their Western peers and are also able to raise significant war chests. The Chinese competitors also have a surfeit of talent, which will not be as able to defect to U.S. tech companies due to Chinese restrictions (see the Manus deal for reference).

The key datapoint to judge the bull markets sustainability are the run-rate revenues of the AI labs. They seem to update the market frequently on various podcasts. If that stays robust, the bull market will likely continue; vice versa if they don’t.

The Marks Are Not Accurate. HG, a software private equity firm, valued Visma, a software company, at 20x EBITDA when public financial comparables are trading at 10x forward earnings. Visma was supposed to be “London’s listing of the decade,” but HG has pulled the IPO. HG hasn’t completed an IPO in over a decade, shuffling assets between funds and using fund-level debt to return cash without real costs. The 2x valuation gap between private and public markets suggests widespread overvaluation in PE-backed software.

IRS Boosts Consumer Spending. U.S. taxpayers received their tax refund checks and may have been spending pre-emptively before energy price increases. The inflationary impact of energy prices should flow through the economy soon. Right before the mid-terms perhaps.

Who Will Rule the UK. Local elections results showed a sharp pivot to Nigel Farage’s Reform Party. Other opposition parties, the Liberal Democrats and Greens, also acquired more seats. But Labour and the Conservatives lost significant seats. There have been calls for Starmer to resign. One data point is that Reform’s popular vote seemed to decline from the last election. The UK faces a challenging fiscal picture as a future government is unlikely to acquire a majority and it’s hard for coalition governments to satisfy bond investors.

China Still Has Property Woes. Property clearance times have hit all-time highs according to SouFun-CREIS data. In Tier I cities it takes an average of 44 months to sell a property, and 65 months in Tier 2 cities. In 2021, it was about 15-20 months. House prices have fallen for 50 consecutive months. Millions of mortgages are underwater prompting banks to extend loan terms, offer interest-only periods and slow foreclosures. The China government can suppress the crisis through administrative means including forced rollovers and directed lending. But it shows the Chinese economy is still quite fragile and very dependent on export revenues.

End Note

If you are in America, happy Mother’s Day! If some of the bouquets seem a little lighter than normal, according to the Groundwork Collective flower prices are up 16% from a year ago driven by tariffs on imported flowers and higher air-freight costs. But it’s the thought that counts.

Gamestop made a non-binding offer for eBay. I asked Claude about the credibility of the bid, and it had the same scepticism as most of the market. “The idea that Gamestop would fund a $56 billion acquisition by selling stocks and memorabilia on eBay is ridiculous. A $20 billion ‘highly confident letter’ from TD Bank for this deal… is not how corporate M&A works.”

TACO has now given way to NACHO (Not a Chance Hormuz Opens). For a President who seems anti-Mexican, he’s certainly embracing their cuisine. And we better hope Hormuz does indeed open.

Omar Sayed