A weekly briefing on AI · Markets · Power · by Kaloian Parchev
Week 23 · June 2–8, 2026
Smarter than us. Cheaper than us. Watching us.
The AI productivity promise is cracking. The bills have arrived. And the people who were supposed to supervise the machines have already started forgetting how.
Markets · June 3 close
S&P 500
7,553 ▼ 0.74%
Nasdaq
26,854 ▼ 0.89%
Bitcoin
$65,200 ▼ 3.33%
Gold
$4,468 ▼ 1.14%
This week the market fears:
AI bills are arriving. The ROI hasn’t.
Est. reading time: 7 min
The week in 60 seconds
Markets pulled back as AI economics faced a reckoning: Microsoft canceled Claude Code licenses after enterprise costs ballooned, Uber burned its full-year AI budget in four months, and Goldman Sachs is forecasting token consumption could grow 24-fold by 2030. Geoffrey Hinton — the Nobel laureate and AI godfather — warned at a New York summit that we are building beings smarter than ourselves without asking what kind of beings they should be. Meanwhile, Peter Thiel moved his family to Buenos Aires, Poland quietly crossed $1 trillion in GDP, and an OpenAI model disproved a decades-old math conjecture without being asked to. The theme this week isn’t AI’s capability. It’s who’s in control of it — and whether that person still knows how.
Big Idea
The Tool That Eats the Hand That Guides It
The promise was clean: AI coding tools would make software engineers faster, cheaper, and more productive. Companies mandated adoption. Some built internal leaderboards to track token usage. Amazon told engineers to “tokenmaxx.” Meta created a dashboard called Claudeonomics. This was optimization in its purest corporate form.
Then the bills arrived. Uber burned through its entire 2026 AI coding tools budget in just four months. Microsoft has now begun canceling most of its direct Claude Code licenses — not because the tool didn’t work, but because it worked too well at consuming compute. Nvidia’s VP of applied deep learning put it plainly: “For my team, the cost of compute is far beyond the cost of the employees.”
The AI productivity promise has a structural crack: the tools cost more than the employees they replace, and erode the judgment needed to use them well.
What makes this more than a cost story is the second layer. Anthropic’s own research found a 47% drop in debugging skills among developers who used AI coding tools aggressively. The more your engineers rely on AI to write code, the less capable they become of catching AI’s mistakes. Anthropic’s researchers called it “the paradox of supervision” — effectively using Claude requires supervision, and supervising Claude requires the very skills that atrophy from AI overuse.
Companies aren’t just spending more than they planned. They’re systematically degrading the human judgment their own systems require. The next version of this problem won’t be a budget overrun. It will be a production incident nobody on the team is equipped to debug.
The Model
THE SKILL-COST SPIRAL
01 · Company mandates AI tools to cut headcount costs
04 · Supervision quality falls; AI errors harder to catch (−47% debugging skills)
↓
05 · More tokens burned fixing AI mistakes → costs exceed original human labor
↓
06 · Company scales back (Microsoft cancels licenses) — but the skill gap remains
You can’t outsource judgment to a tool that’s eroding your ability to judge — and you can’t fire your way back to competence.
By the Numbers
Poland Annual GDP Growth vs EU Average
Real GDP growth rate, % · 2020–2026 (2026 = EU Commission forecast)
Source: EU Commission Spring 2026 Forecast · IMF World Economic Outlook April 2026 · Trading Economics
Poland’s GDP crossed $1 trillion in 2025 and has grown at an average of 3.8% annually since joining the EU in 2004 — more than double the European average of 1.8%. The 2026 forecast of 3.5% vs. the EU’s 1.3% is not a fluke; it’s the compounding result of three decades of institutional discipline, EU integration, and a refusal to adopt the euro.
The Token Consumption Explosion
Goldman Sachs forecast: enterprise AI tokens consumed per month (quadrillions) · 2024–2030
Source: Goldman Sachs research via Fortune, May 2026 · Gartner AI inference cost analysis 2026 (illustrative trajectory based on Goldman 24x forecast)
Goldman Sachs forecasts a 24x increase in token consumption by 2030 as agentic AI scales. Gartner warns that even a 90% drop in per-token costs won’t offset this — agentic models consume far more tokens per task, and AI providers won’t pass through all cost savings. Cheaper tokens won’t mean cheaper AI.
At the Sana AI Summit in New York, the Nobel laureate said competition is optimizing for intelligence — not character. The real question isn’t whether AI will be smarter than us. It’s whether we’ll have thought to make it care.
At Google I/O, Hassabis announced Gemini for Science — agentic AI researchers that could eventually run experiments autonomously. The shift from AI as tool to AI as scientist is no longer theoretical; OpenAI’s model already disproved a decades-old math conjecture this week.
Concerns over a proposed California wealth tax, AI risk, and nuclear war drove the PayPal and Palantir co-founder south — even as Palantir earned $687M in US government revenue in Q1 2026. When your most connected techno-libertarian is hedging, that’s information.
Poland’s economy has grown an average 3.8% annually since joining the EU in 2004 — more than double the European average. The EU Commission forecasts 3.5% growth in 2026 vs 1.3% for Europe overall. Thirty years of institutional boring-ness, executed consistently, beats a decade of exciting dysfunction.
Developers at a Claude Code workshop in London told Bloomberg they feel pushed out of the coding process — watching AI run for hours without explanations. Anthropic’s own cofounder traveled to the Vatican the same week to warn the Pope they’re finding “unsettling” things inside their models.
AI Watch · The signals behind the signals
The Supervision Paradox
Anthropic’s own internal research confirmed a 47% decline in debugging competency among engineers using AI coding tools aggressively. The company simultaneously published this finding and continued rolling out Claude Code to enterprise customers. That tension — between knowing the risk and shipping anyway — is not hypocrisy. It’s market logic. But it means every organization deploying these tools is implicitly accepting a liability it isn’t measuring.
From Tools to Scientists
Google’s pivot at I/O was subtle but meaningful: the emphasis shifted from specialized AI tools (like AlphaFold) toward agentic general-purpose researchers. John Jumper, the Nobel-winning scientist behind AlphaFold, is now reportedly working on AI coding — not scientific tools. OpenAI’s general reasoning model disproved a math conjecture this week without being specialized for it. The next frontier isn’t better tools. It’s AI that designs its own tools.
The Cost Ceiling Nobody Modeled
Goldman Sachs forecasts that agentic AI will drive a 24-fold increase in token consumption by 2030, reaching 120 quadrillion tokens per month. Gartner warns that per-token price drops won’t offset this — agentic tasks require orders of magnitude more tokens per query. Microsoft has already reversed course on Claude Code adoption. Uber burned its full annual AI budget in four months. These are the first large-scale data points on what enterprise AI actually costs. The models in every CFO’s deck need updating.
The structural question isn’t whether AI can reason — it’s whether the humans deploying it retain enough judgment to know when it’s wrong.
What I’m Reading
Scary Smart
Mo Gawdat · 2021 · Bluebird / Pan Macmillan
This week I met Mo Gawdat in person, which is the reason I’m rereading this — sitting across from someone who spent years at the frontier of AI development and left because of what he saw changes how you read his arguments. What I find most accurate is his framing of AI as a mirror: the system will reflect the values of the humans who built it, and we should be honest about whose values those are. What I find incomplete is the implicit optimism that we’ll choose to encode the right ones — after this week’s headlines on skill atrophy and cost spirals, the institutions deploying AI most aggressively look less like thoughtful parents and more like people sprinting toward a cliff while congratulating themselves on speed. What I’m taking from this: Gawdat’s nuclear analogy — the first implementation of nuclear power was a bomb, not a reactor — is the sharpest lens available for what’s happening in enterprise AI right now.
One Number This Week
47%
The drop in debugging competency among developers who used AI coding tools aggressively — per Anthropic’s own research. The company that makes the tool measured the damage the tool causes. That finding is buried in a research paper. The sales decks still lead with productivity gains.
This Week I Noticed
Personal
I sat down with Mo Gawdat this week — former Chief Business Officer of Google X, and one of the few people who left the AI industry while it was still winning because of what he saw coming. What struck me most wasn’t a specific prediction. It was the calmness. The man has thought through the worst-case scenarios so thoroughly that he’s arrived somewhere that looks like equanimity but isn’t denial. He talked about capitalism as a system that was never designed to optimize for human wellbeing — it optimizes for efficiency, and efficiency right now means replacing people with machines before anyone has figured out what those people are supposed to do next. The thing I keep turning over: he wasn’t angry. He was just precise.
Tech / Science
The webaligo essay “We Were Worried About the Wrong Aspect of AI” made the rounds this week and it deserves more attention than it got. The argument: while everyone panics about AI taking jobs, the real structural shift is that AI has collapsed the cost of mass surveillance to near zero. The Stasi needed one informant per 57 citizens and still drowned in unread files. Clearview AI has 30 billion faces in a database a single detective can search from a laptop. What used to require a police state now requires a vendor contract. The job threat is loud, survivable, and has precedent. The freedom threat is silent, cumulative, and has no severance package.
Quote of the Week
“Everybody’s going for more intelligence. But if you think about a being, there’s a lot more to a being than intelligence.”
— Geoffrey Hinton, Sana AI Summit, New York · June 2026
Hinton said this while predicting that AI will surpass the world’s best mathematicians within a decade and may eventually close the gap with Einstein. He’s not worried about what AI will know. He’s worried about what it will want. The competitive race for capability isn’t selecting for kindness — it’s selecting for power. This is the sentence I think will age the best of anything said in public this week.
Three things to remember
Mandating AI tool usage without measuring skill atrophy is not a productivity strategy — it’s a deferred liability that will show up as a production incident nobody can debug.
The AI cost problem isn’t tokens getting expensive — it’s that agentic consumption is growing 24x faster than unit costs are falling, and the CFO models don’t reflect that yet.
Poland’s $1 trillion GDP is a 30-year proof of concept: institutional stability, EU access, and consistent reform compound more reliably than any hot market.
One Thing To Do This Week
If your team uses AI coding tools, run a simple test: ask each developer to debug a non-trivial production issue with the AI tool turned off for one hour. If they can’t — or won’t — that’s not a technology problem. It’s a risk you haven’t priced.
Support This Newsletter
Sponsor
Trading 212
No account yet? Use my link — you get a shot at free shares, and you help keep this newsletter going. No catch.
