$950B Korea Deal Masks Agent Cost Crisis

So, What Actually Happened?

So I spent Saturday morning adding up the week's numbers and gave up somewhere past nine hundred billion. South Korea signed $950 billion in AI deals off the back of its president's Washington trip, and Samsung won a $200 billion Broadcom partnership for its foundries. We scanned 190,000 articles this week so you don't have to, and the loudest thing in the corpus was concrete: fabs, racks, power. Then I noticed what people were actually arguing about. Not capacity. Copying. The whole industry spent the week obsessed with distillation, the trick that makes an expensive model cheap. And McKinsey quietly reported that 60% of agentic AI spend goes to re-running answers until they are good enough to ship.

The Bottom Line: The concrete is being poured at nation-state scale while the margin gets attacked from underneath, by a copying technique and a cost line nobody budgeted for.

The Tracks That Matter

1. South Korea Signs $950 Billion Into the AI Supply Chain

Presidential trips usually produce photographs. This one produced a supply chain. Following the South Korean president's Washington visit, Samsung and SK Group sealed $950 billion in AI commitments, with SK Group carrying roughly $750 billion of it, including an SK Hynix arrangement with Nvidia valued above $500 billion. Those are not procurement numbers. Those are budget-of-a-mid-sized-country numbers. Nvidia also put $1 billion into Naver to anchor Korea's sovereign AI stack, which tells you the shape of the deal: memory, foundry and power locked in under national signature rather than commercial contract. Compute just became a line item in a trade negotiation.

Here's what works: Ask your cloud team which AI workloads now sit on capacity contracted at government level. Those prices move on politics, not on your renewal date.

2. Everyone Suddenly Cares How Cheaply You Can Copy a Model

The week's real fight was not about building. It was about copying. Distillation, training a small cheap model on the outputs of a big expensive one, went from research footnote to the argument running from Silicon Valley to Washington, and the lawyers arrived with it: whether the practice is even legal is now an open question. The US Treasury is weighing sanctions and export curbs on Chinese labs over exactly this. Meanwhile the British and American safety institutes found that Kimi K3's safeguards failed to block offensive cyber attempts, which is what happens when capability arrives cheaply and the guardrails do not travel with it. Cheap copies are the competitive threat and the security surface at the same time.

Here's what works: If you fine-tuned on another vendor's outputs, find out this week whether your contract allowed it. That answer is about to get expensive.

3. Most Enterprises Are Already Over Budget on Their Agents

The number that should have led every wire this week did not. McKinsey found that 60% of agentic AI costs go to response refinement, the loop where a system re-runs and reworks its own output until it is good enough to ship, and that most enterprises have already blown their budgets. Not on licences. On retries. Public markets are noticing the same thing from the other end, with the trillion-dollar infrastructure bet showing its first cracks as investors go hunting for profit elsewhere. Everyone modelled agents as labour substitution. The real cost curve looks more like cloud egress in 2015: invisible in the pilot, brutal in production, and owned by nobody.

Here's what works: Ask your platform team for cost per resolved task, not cost per token. If nobody can produce that number, your agent budget is a guess.

Quick hits:

• The agent market started buying itself. Sierra acquired Takeoff to build long-horizon agents, one of several moves as the market enters consolidation — the pilot phase is closing, the shopping phase has opened.
• A chip startup raised $300M to do one thing well. Etched closed a $300 million Series C for inference silicon, a direct bet that running models, not training them, is where the money now sits.
• DeepSeek stopped taking money. The Chinese lab paused fundraising talks with prospective investors, an odd move in the hottest capital market in tech unless you no longer need the capital.

Signal vs. Noise

🟢 Signal: Inference economics. Etched raised $300 million for chips that do one job, distillation became the industry's obsession, and McKinsey found most enterprise agent programmes already over budget on retries. Three unrelated desks, one shared assumption: the cost of running a model is now the battleground. Most coverage is still busy counting announcement dollars.

🔴 Noise: ”Agentic AI” as a label. It remains one of the loudest phrases in the corpus and it keeps losing its grip on the actual conversation, which has moved to cost per resolved task and who signs off on the retry bill. If a vendor deck says ”agentic” more often than it says ”per-task cost”, you are reading a 2025 pitch.

Try It Yourself

From the 190K

We scanned 190,000 articles this week. Here's what no one's talking about:

South Korea committed $950 billion to AI hardware, Washington started weighing sanctions over model distillation, and McKinsey reported most enterprise agent programmes are already over budget, all inside the same 48 hours.

Each of those lands on a different desk. The trade desk writes up Korea. The policy desk writes up the sanctions. The enterprise-software desk buries the McKinsey finding in a Tuesday briefing nobody reads. Put them on the same morning and they are one story: the industry is pouring an enormous, capital-intensive floor under a product whose economics are being squeezed from two directions at once. Distillation attacks the price of capability from above, by making the expensive model cheap to imitate. Retry costs attack the margin from below, by making each delivered answer cost more than anyone modelled. The $950 billion is a bet that scale arrives before either squeeze closes.

What changes on Monday is much smaller than any of that. Take the three AI systems your business genuinely depends on and ask, for each one, what a single completed piece of work costs you today. Not the licence. Not the token price. The delivered unit. Most teams cannot answer, and that is the number the next budget cycle gets fought over.

By The Numbers

• South Korea sealed $950 billion in AI deals with Samsung, SK and US tech firms — signed off the back of a single presidential trip to Washington.
• Samsung won a $200 billion Broadcom AI chip partnership — a foundry order large enough to reset the contract-manufacturing hierarchy.
• SK Group and Nvidia announced a $500 billion AI infrastructure partnership — memory and compute tied together at a scale no single enterprise buyer can influence.
• Etched raised $300 million in Series C funding for AI inference chips — private capital voting for cheaper running costs over bigger models.
• OpenAI is targeting an $830 billion-plus IPO valuation but will not break even until 2028 or 2029 — the gap between what the market prices and what the business earns, in one sentence.
• See what's rising in our 190K-article corpus this week →

Deep Dive: Everybody Announces the Tour, Nobody Announces the Trucking Bill

I promoted a few shows in my twenties. The announcement is the easy part — big venue, big name, big number on the poster. What kills you is everything that never makes the poster: the trucks, the crew, the rider, the soundcheck that runs into overtime. This week the AI industry announced the tour.

The venue deals got signed
South Korea put close to a trillion dollars of memory, foundry and infrastructure under national signature. Samsung's foundry business is back in the top conversation. SK Hynix is bound to Nvidia at a scale that reads like fiscal policy rather than a supply agreement. That is the poster: enormous, credible, entirely about capacity.

The support act is undercutting the headliner
Then distillation went mainstream. Train the small model on the big model's answers and you get most of the show at a fraction of the production cost. Washington's response has been legal rather than technical, because there isn't a technical response available. Once the copy is good enough, the moat stops being the model.

The rider is where the money actually goes
And inside the enterprise, the bill landed. Most of what companies spend on agents is not the answer — it is the retries. Evolution, not revolution: the teams that come through this are the ones already measuring what one completed piece of work costs them, before the CFO thinks to ask.

What Actually Works

1. Measure cost per resolved task: Tokens are an input price. A completed piece of work is the unit finance will actually ask about.
2. Audit what you trained on: If you fine-tuned on another provider's outputs, get the contract position now, not after the first letter arrives.
3. Name the supply-chain politics: Any workload sitting on government-contracted capacity carries geopolitical price risk. Write it down where procurement can see it.
4. Cap the retry loop: Set a hard ceiling on refinement passes per task and log what falls off the edge. The failures teach you more than the successes do.

The poster sells the tour. The trucking bill decides whether you go home with money.

What's Coming

Distillation Becomes a Legal Category

Whether copying a model's behaviour is legal remains an open question with commercial answers pending. Within two quarters, expect model licences to carry explicit anti-distillation clauses and expect legal teams to start asking what your fine-tuning data was derived from. There is no good improvised answer to that question.

Sovereign Compute Turns Into a Procurement Geography

The buildouts announced this week have to land somewhere physical, and power and permitting are already binding on data centre investment. Expect ”where does this capacity sit, and under whose law” to appear on vendor questionnaires alongside SOC 2 before the year ends.

Agent Contracts Get a Per-Outcome Line

Once enterprises are over budget on refinement, finance stops accepting token-based forecasts. The next generation of agent contracts will price per completed outcome, and vendors who cannot commit to that number will lose renewals to the ones who can.

For Your Team

Monday's meeting prompt: ”If we had to state the cost of one completed piece of AI-delivered work today, could we? And if not, who owns finding that number before the next budget cycle starts?”

Share-worthy stat: South Korea signed $950 billion in AI deals in a single week, on the back of one presidential trip, at the same moment McKinsey found most enterprises already over budget on the agents that capacity is meant to serve. The concrete is running ahead of the economics.

Go deeper: Track where AI cost pressure is really moving, in real time →

The Track of the Day

”IPO hype now centers on which companies can prove unit economics, not which has the biggest model.”
From a market analysis published this week

Two years ago that sentence would have read as a loser's argument, the thing you say when your model isn't the biggest. This week it reads like an entire industry catching up to its own invoice.

We scanned 190,000 articles this week so you don't have to. Data Pains → Business Gains.

Published: July 26, 2026 | Curated by Yves Mulkers @ Ins7ghts