본문 바로가기
IT Strategy Weekly

Nov 8 – Nov 14, 2026

Korean original

Not investment advice — these are resources for learning and forming your own view.

The real shift this week was not the race for model performance. The source of AI’s value moved from ‘who holds the smartest model’ to ‘who is free of the physical and political choke points of power, chips and regulation’. Companies that have handed that control to others lose bargaining power, however good their models are.

The strategist's view

In AI investment due diligence, give 'model performance' less weight and move 'grid connection timeline' and 'export-control clauses' to the top of the list.

Read this week's news as one story and it shows that AI's bottleneck has moved twice. First the bottleneck was the model, then the chip. Now it is the electricity to run those chips and the political permission to use them. In the same week, one analysis found that data centres cannot be built because they are stuck on grid access and permits, and a report said Nvidia had confirmed to the government that its chips can be location-tracked. That is not a coincidence; they are two faces of the same structure.

If so, the investment committee's questionnaire has to change too. Until now we have asked 'which model will we use?' and 'how many GPUs have we secured?'. The answers to both have converged, so they no longer tell options apart. What we should ask instead is: when is the confirmed grid connection date in this contract? Who bears the loss if a regulatory change halts operations? And are the data and evaluation criteria we are building assets that come with us if we switch suppliers?

The trend of half of vertical AI companies starting to train their own models follows the same logic. They are not building models to show off performance. They know that if they stake the core of their business on someone else's model, everything is shaken when prices and policies change. A company that has not decided explicitly how much control to hold itself and where to start borrowing has not avoided the decision. It has decided to borrow everything.

Cases through a framework 3

Vendor Lock-in

Nvidia privately confirmed it could track chips to fight smuggling, even as industry publicly questioned feasibility

Nvidia privately confirmed that it can track the location of advanced AI chips, in line with the requirements of the Chip Security Act under discussion in Congress. In the same week came reports that export-controlled GPUs were used to train China's top models, along with the CEO resignation affair. US politicians have put Nvidia's China sales under investigation.

Buyers' AI stacks are tied to one supplier's proprietary technology — CUDA, dedicated interconnects and the operating data built on top of them. Now government conditions, location tracking and export controls, are being added to those chips. A legal barrier sits on top of the technical barrier that already blocks switching: a textbook lock-in structure.

This news is usually read as 'a US security measure to stop smuggling into China'. Through the lens of lock-in, though, the point is not China. Buyers worldwide who boarded a single supplier for its early performance and convenience now find that the operating conditions of their assets change with policy shifts in Washington, not with the supplier. The real cost of lock-in comes not from the contract term but from a structure you cannot exit.

Decision prompt — Starting with next quarter's GPU procurement contracts, always include a clause that states who bears downtime and region-relocation costs if export-control or tracking requirements change. And commission now, with a six-month deadline, a redundancy test that runs at least 20–30 percent of inference workloads on alternative accelerators or in regions under other jurisdictions.

About the framework

Once you are deeply tied to a supplier's technology, it is hard to leave even when it raises prices or changes terms, because proprietary formats, dedicated APIs and accumulated data make switching practically impossible. So a low early price or early convenience comes back later as lost bargaining power. The real cost of lock-in comes not from the contract term but from a structure you cannot exit.

MIT Technology Review

Jobs to Be Done (JTBD)

14 of the 30 most valuable vertical AI companies now train their own models

Of the 30 most valuable vertical AI companies, 14 have moved beyond building on general-purpose models and now train their own, building lab-level capability in-house. In the same week OpenAI acquired the travel start-up Odessia, moving inside the workflows of a specific domain.

These companies' customers are not buying 'a good model'. They pay to get one narrow, repetitive chunk of work done, such as insurance claims review or clinical documentation. General-purpose model APIs, in-house staff and existing SaaS compete across categories for that work, so the scope of the 'job' becomes the unit of analysis.

This news is usually read as 'vertical players now build models too, colliding head-on with foundation model companies'. Through the jobs-to-be-done lens, what they are building is not a smarter model but a complete capability that brings the failure rate on a specific task below human level. So the contest is decided by rework and approval rates on that task, not by benchmark scores. Conversely, OpenAI's domain acquisition is also an admission that a general-purpose model cannot finish that job on its own.

Decision prompt — Do not start by asking whether to train your own model. Pick three tasks tied directly to revenue and, within two quarters, build an evaluation set that measures each by failure and rework rates, plus a proprietary data pipeline. Spend in-house training budget only on tasks where general-purpose models ultimately fail to reach the target level on that evaluation set.

About the framework

People do not buy products; they hire them to get a 'job' done in their lives. So the true scope of competition is not products in the same category but every alternative that could do that job. However far you raise the specs, a product that misses the job customers actually want done will not sell, and if an entirely different technology does the job better, the whole category is replaced.

CB Insights

Technology S-Curve

Why the data center you dread might never even be built

The AI data-centre building boom is in fact slowing, held back by local opposition, permitting delays and power shortages, and one analysis argued that public opposition to data centres is really about grid overload, not corporate power. In the same week X-energy received a construction permit for a small modular reactor, and Fervo secured USD 12 billion in geothermal orders.

We have entered a phase where adding inputs (capital and land) no longer raises compute capacity as before, because of the physical limits of grid-connection queues and permits. At the same time, entirely different power curves — geothermal, small reactors, floating offshore — are rising from below into the commercial-order stage. The conditions for a curve switch are in place.

This news is usually read as 'a sign the AI bubble is bursting' or 'NIMBYism is blocking innovation'. Through the S-curve lens, what is flattening is not demand but the performance curve of the existing way of procuring power. The contest is no longer about securing a few more GPUs. It has moved to who first takes a position on the next curve: self-generation and new baseload power.

Decision prompt — Rewrite new data-centre contracts around the confirmed grid-connection date rather than the completion date, and attach compensation clauses for delays. This quarter, revise investment approval criteria so that advance payments for capacity added from 2027 go only to sites backed by geothermal or small-reactor power purchase agreements.

About the framework

A technology's performance relative to input traces an S: slow at first, explosive in the middle, flattening again at maturity, because the closer it gets to physical and structural limits, the less extra investment adds. So as one technology nears the top of its curve, an entirely new technology's curve rises from below and eventually takes over.

TechCrunch

Based on 108 items over 7 days

Past issues