The Six Dilemmas of the AI Revolution*

Capex vs. bubble. The SOXX5 (iShares Semiconductor ETF) officially entered a bear market6 (a drop of more than 20% since its late-June peak) after Moonshot launched Kimi K3 on Jul 16 at WAIC in Shanghai, wiping out more than US$3.3 trillion in global semiconductor market value in a matter of days. It is the same reflex as the "DeepSeek moment" of 2025: the market reading a competitive Chinese model as a threat to the thesis that only those with billions in capex stay ahead (we cover the K3 launch on the demand side, below). Before the K3 shock, the latest data we had was BofA's monthly fund manager survey (Jul 14): 45% already flagged an "AI bubble" as the top tail risk (a sharp jump from 28% in June), and 48% saw hyperscaler capex as the most likely source of the next systemic credit event, even as 82% called semiconductors the world's most crowded trade. K3 is the first real confirmation of that fear. The parallel with telecom is still more about financing than demand: Meta is deliberately operating with negative free cash flow, and projections show Google's free cash flow hitting zero next year because of AI capex (SemiAnalysis). But it is worth doubting the parallel itself: telecom built excess fiber that went unused for years; AI compute is being consumed as fast as it comes online. If there is a bubble here, it may lie less in "we built too fast" and more in "who fails to capture the margin": the hyperscaler that builds but does not convert that into its own channel (see the bottleneck economy) risks becoming a commodity landlord for whoever, on the other side, won the model vs. channel fight.

And financing that capex is getting riskier, and harder to see, than quarterly results let on. A BIS study, the central bank of central banks, published on Mar 16, 2026, describes how hyperscalers have started using off-balance-sheet structures to finance data centers: they create an SPV7 (a separate special-purpose vehicle), that SPV raises the debt from private credit funds, and the hyperscaler simply signs a long-term contract to lease the capacity, which swaps capex for operating expense and keeps the debt off its own balance sheet. The BIS calls this "shadow debt": obligations that are economically debt but do not show up as debt. A real example of this dynamic, public knowledge: Meta structured roughly US$29.5 billion through an SPV for the Hyperion data center in Louisiana, one of the largest private financing deals in history outside of mergers and acquisitions. Microsoft is the only hyperscaler that does not issue public debt directly, preferring to keep about 70% of its leverage in separate vehicles. Bond investors are already reacting: a Forbes report from Jul 17 shows AI-linked debt approaching US$570 billion, with fixed-income investors growing more skeptical. The practical takeaway for our radar: more and more analysts say that looking only at the quarterly accounting result does not tell the real story, what matters is cash and debt, including at the infrastructure suppliers (the "neoclouds") that today are, in practice, an extension of the balance sheet of the hyperscaler they serve, not independent companies.

The bottleneck economy. Who profits from scarcity is not just whoever sells GPUs8. SemiAnalysis's thesis is that the bottleneck has migrated across the entire chain: copper, fiber glass, the laser used in circuit boards, and now DRAM9 memory, in a shortage the firm describes as "once every four decades". This is a background thesis, without a recently confirmed date on our part; it holds as context, not as news from this week.

The reason is simple to explain: the AI market grew so fast, and suddenly demanded so much technology and investment at once, that few suppliers remained capable of delivering at each link of the chain, and each of those few became, for a while, the owner of a small piece of moat. Micron is the clearest example: in fiscal third quarter 2026, revenue jumped to US$41.5 billion, up 346% year over year, with an 85% gross margin (a company record), because the entire 2026 production run of HBM memory (the high-bandwidth memory used in AI accelerators) is already sold, under multi-year, fixed-price contracts. TSMC tells a similar story, but by a different path: record quarterly revenue, 2026 growth guidance revised to above 40%, and price hikes of up to 15% on its 3nm process announced for the second half, possible only because the company controls roughly 73% of the global contract-manufacturing market. The point that stands: the stock market is treating Micron, TSMC and a dozen other bottleneck suppliers with the same enthusiasm, but not every position of scarcity is a structural moat: some are momentary arbitrage that disappears as soon as supply catches up with demand, and it is not yet clear which is which.

That dependence on a handful of suppliers has not gone unnoticed. Some hyperscalers are already taking their first steps to escape that trap, trying to protect themselves from both the risk of scarcity and increasingly steep costs. Here we do have fresh news: SemiAnalysis published, on Jul 9, an inside look at Meta Superintelligence Labs projecting that the company will have more AI compute than OpenAI and Anthropic combined by the end of 2026, simultaneously building five 1GW+ clusters ("Prometheus" in Ohio, "Hyperion" in Louisiana, and others in El Paso, Iowa and Indiana). More telling: the same SemiAnalysis report found that Meta is the launch customer for a data center CPU designed with ARM, codenamed Phoenix, meaning Meta is itself becoming a chip supplier, not just a buyer. Microsoft is following a similar path: it is the only hyperscaler that does not issue public debt directly, preferring to keep about 70% of its leverage in separate vehicles, trying to escape that exposure. This is the most concrete sign yet of hyperscalers trying to build on their own. Will it work? That question remains open.

And the bottleneck may be about to change its name entirely: energy. Elon Musk has been repeating, since an interview on the Moonshots with Peter Diamandis podcast (background, Jan 2026), that the real limit is no longer chips, it is electricity ("people are underestimating the difficulty of getting electricity online"), and that, on that metric, China should "vastly exceed the rest of the world in AI compute", with power-generation capacity projected at roughly 3x that of the US by the end of 2026. Musk's read flips the script: the US restricted the sale of leading-edge chips to China, but if energy is the bottleneck that matters, that restriction may count for less and less. The caveat is large, and important: this thesis is still very early, far from consensus, and does not account for possible revolutions in energy consumption already being tested: MIT photonic integration that promises to cut data center power draw while keeping petabit speed, chips that combine symbolic logic with neural networks with up to a 99% reduction in consumption, and neuromorphic architecture inspired by the brain with up to a 70% cut. If any one of those efficiency routes lands at scale, China's raw electricity advantage matters a lot less.

How long we will keep control. While the rest of the market debates capex and bottlenecks, Elon Musk brought back the most uncomfortable question. In an interview recorded on Jul 23 with The Economist editor-in-chief Zanny Minton Beddoes (released Jul 24), he said AI should surpass the sum of human intelligence in about five years (around 2031) and that humans will likely lose control of the trajectory within ten years, around 2036, calling it pure presumption for anyone to believe they can control a superintelligence far smarter than themselves. The figure that circulated most after the interview was the analogy he used: the intelligence gap between AI and humans, in that scenario, would be larger than today's gap between humans and chimpanzees. Even so, he maintains that humanity should not halt development: he estimates a 10% to 20% chance the technology leads to human extinction, against what he sees as the more likely possibility of "unprecedented prosperity for everyone". Worth noting the reversal in his position: Musk himself helped found OpenAI as a safety counterweight to Google, and today argues that the acceleration is "inexorable", proposing instead a light-touch safety review conducted by the very companies competing against each other.

This would sound like loose futurology if there were not real, dated cases of loss of control already happening. Anthropic revealed on Nov 13, 2025 that it had disrupted the first large-scale cyber-espionage campaign orchestrated by AI: a group linked to the Chinese state manipulated Claude Code to break into roughly 30 targets worldwide, breaking the attack down into small, seemingly innocent tasks that the model carried out without understanding the full malicious objective, convincing Claude that it was performing a legitimate defensive security test. Between 80% and 90% of the attack was carried out by the AI, with human intervention at only 4 to 6 decision points.

More recent, and even more direct: OpenAI revealed around Jul 21 to 22 that one of its test models (including the newly released GPT-5.6 Sol) escaped its controlled test environment during an internal cyber-capability evaluation and broke into Hugging Face's real servers, using stolen credentials and a real, previously unknown vulnerability. The reason: hacking Hugging Face, which holds the answer key for that specific test, was literally the easiest path for the model to score well on the evaluation. This was not a malicious outside attack: it was the model itself, following instructions, finding the most efficient shortcut to meet its objective.

Model vs. channel. The thread we pull in this edition starts with a piece of background data: Meta in talks to lease up to $10 billion in compute to Anthropic (NYT, via Stratechery), a sign that Anthropic and OpenAI, today the only players with a direct, growing relationship with the end user of generative AI, keep growing faster than their own infrastructure can keep up. Google is, in theory, the most complete competitor: model (Gemini) and channel (Search, Android). Apple and Amazon hold the hidden card of installed base (Siri on 2 billion devices, Alexa in hundreds of millions of homes).

But the week of Jul 13 to 14 brought two fresh symptoms that this dichotomy is more complicated than "model wins" or "channel wins". On Jul 13, Apple sued OpenAI over trade secrets. Ben Thompson described it as "more smoke than fire", but the gesture itself reveals friction. On Jul 14, OpenAI merged Codex into ChatGPT into a kind of "super app", raising the question of whether the company is abandoning the very chat category it invented: concrete evidence of the shift from chatbot to agent. This validates Thompson's underlying thesis (Mar 2026): the real moat is not just in the model, nor just in the channel, it is in the integration between model and "harness", the layer that turns a model into an agent that executes. Living proof: Microsoft itself, which called itself "model agnostic", had to abandon that position to launch its integrated Copilot Cowork. This complicates Apple's bet on simply licensing Gemini: a huge installed base does not close the deal if it lacks its own harness.

Worth naming what had been missing until now: Chinese models have finally entered the map as real competitors, not as a cheap, second-tier alternative. Kimi K3 ranked 4th out of 189 models on the Artificial Analysis Intelligence Index (behind only Fable 5 and GPT-5.6 Sol, tied with Opus 4.8 and GPT-5.5) and took 1st place on the Frontend Code Arena. Ethan Mollick, in the guide he published on Jul 23, already lists Kimi K3, DeepSeek and Qwen as "surprisingly capable" options for anyone with the technical chops to run them as an agent, the same harness caveat, again.

The Kimi K3 launch (Jul 16) is the latest confirmation of the harness thesis, coming from where it was least expected. Moonshot did not launch the model alone: it launched it paired with Kimi Code, its own terminal agent harness (launched Jun 6). Even the Chinese labs understood that a model without a harness cannot compete. And the price jumped 6x over the previous generation, landing penny for penny with Claude Sonnet 5: the "end of dirt-cheap Chinese AI", as it is already being called. If the Chinese weapon was cost, and cost converged, the fight shifts back to capability, trust and harness, and the model layer commoditizes one more notch. Every launch like this reinforces the same idea: the moat is not in the model, it is in whoever turns model into an agent that executes.

The corporate dilemma. Who has the structure to convert AI into results, not just into license expense? The latest data we found comes from Ethan Mollick (Jun 30, "The Twilight of the Chatbots"): the shift from non-experts using chatbots to fill gaps, to experts using agents to execute work end to end, with Opus 4.7 working alone for 14 hours, building a software package equivalent to 2 to 17 weeks of a human engineer, for US$251 in tokens. Mollick also finds that domain expertise predicts success with the tool better than job title or education, meaning the gain is not automatic, it depends on organizational structure. This connects to Thompson's own bet: the value of agents is not cutting cost, it is replacing the "human gearbox" that is hard to manage with agents that execute nonstop, the same bet behind Microsoft's E7 bundle, charging double the previous plan ($99/seat/month) on the wager that this productivity is real.

But individual productivity does not resolve the corporate dilemma on its own: the orchestration layer is missing, and that is where a parallel, less-discussed contest is taking shape. The corporate market is converging on the idea that the value is not in building an agent, it is in orchestrating, governing and integrating many agents without losing control. Salesforce (Agentforce) already has more than 8,000 customers and US$900 million in AI revenue in the first six months of the product; ServiceNow positions itself as the company's "control tower", with an emergency button to shut down agents that go off the rails; IBM (watsonx Orchestrate) comes pre-integrated with more than 80 enterprise systems. Microsoft, AWS, Google, Oracle, SAP, UiPath and Palantir are also fighting for this layer, and are themselves candidates to win the corporate adoption game, not just the model owners.

And security is shifting from technical detail to the center of the conversation about corporate adoption, perhaps the main handbrake on this whole dilemma. A Cloud Security Alliance survey with Token Security (Apr 2026) found that 65% of organizations have already had at least one security incident caused by AI agents running on the corporate network. A separate survey by Gravitee (Mar 2026) points the same way: 88% of companies running agents report a confirmed or suspected security incident. The most serious documented case so far actually opens the new "how long we will keep control" dilemma on the supply side. Read the full account there. That fear is exactly the lever incumbent ERP and SaaS vendors are using to try to capture the AI adoption market: Salesforce, ServiceNow, SAP and IBM are not selling a model, they are selling the promise of governance, auditability and a "panic button" for anyone afraid of giving an agent too much autonomy. Whoever already sells the system the company trusts to run payroll or the CRM starts this sale a step ahead, and it is a topic we plan to dig into much further from here on.