Sep 6 – Sep 12, 2026
English translation of the Korean original, prepared with AI assistance. Korean original
Not investment advice — these are resources for learning and forming your own view.
The strategist's view
Models leak, electricity doesn't -- in 2026, the AI moat lives in power contracts and data ownership documents, not code.
Thread this week's news together and it reads like this: Anthropic disclosing distillation attacks, with government agencies now stepping in too, amounts to something close to an admission that the imitation barrier around model capability was thinner than assumed. On the other side, gigawatt-scale load swings are shaking power grids, and state governments are tightening development with clean-power standards, making electricity and land ever harder to secure. One kind of asset gets copied with an API call; the other is obtainable only through permits and five-year contracts -- yet investment allocation still tilts toward the former.
So my recommendation is to change the question asked in executive meetings. 'How much better is our model than the competition' is a question with a six-month half-life, but 'how many years would it take to replicate the power, land and proprietary data we've secured' directly determines how long the business lasts. Google opening up TPUs externally and Nvidia loading ecosystem demand onto its own books through $530 billion in guarantees both reflect the same judgment: the more commoditized raw compute becomes, the more valuable the physical rights to run it get.
What needs to happen right now boils down to three things: measure inference workload vendor-switching costs in hard numbers and convert them into negotiating leverage; re-order data center specs against next-generation power architecture standards; and shift budgets from revenue plans premised on a model moat to plans grounded in physical assets and data ownership. While everyone else compares model benchmarks, whoever locks up generation equipment behind the meter and data licensing first will win on cost structure a year from now.
Cases through a framework 3
Powering AI is an architecture problem
Analysis emerged showing that gigawatt-scale AI data centers create load swings of 70% within milliseconds, which existing low-voltage power stacks and local grids are structurally unable to withstand. In the same week, reports showed hyperscalers treating behind-the-meter (BTM) self-generation strategies as effectively a condition of survival, alongside analysis that even portable power devices have hit the physical ceiling of passive heat dissipation.
The power stack is a domain where performance per unit of investment can be measured very precisely, in watts, voltage and heat, and the existing approach -- low-voltage distribution and passive heat dissipation -- has already entered a saturation zone where more investment yields little more performance, while a different curve is rising underneath it: high-voltage DC distribution, self-generation and active cooling. Massachusetts attaching new clean-power standards to data centers is also institutionally pressing down on the top of the old curve, accelerating the shift to the new one.
This news is often read as a supply-shortage story: 'generation capacity is short, so build more power plants.' But through the S-curve, the problem isn't the quantity of capacity -- it's the lifespan of the architecture. A data center built to today's standard design will already be standing on an aging curve by the time it's finished, and an operator who mistimes the switch to the new curve ends up absorbing a loss where the asset's useful life is shorter than its payback period.
Decision prompt — Scrap the practice of defaulting new data center specs to low-voltage stacks and passive cooling this quarter, and re-issue them with high-voltage DC distribution, liquid cooling and battery storage to absorb sharp load swings as design premises. At the same time, re-score site candidates not by grid interconnection queue position but by whether they can meet both BTM self-generation and clean-power requirements, and lock in power contracts on long-term fixed terms within six months.
About the framework
A technology's performance relative to investment traces an S-shape -- slow at first, explosive in the middle, and flattening again at maturity -- because as it nears physical and structural limits, performance stops improving no matter how much more is invested. So around the time one technology reaches the top of its curve, an entirely different new technology's curve rises up from underneath and eventually takes over.
MIT Technology Review
TPU Inference Externalization Full Steam Ahead - InferenceX
Google is opening TPUs, long kept for internal use only, to the outside inference market in earnest, building infrastructure that it claims has an edge in cost-performance. In the same week, Nvidia's 70% growth forecast came alongside the revelation of a $530 billion off-balance-sheet guarantee structure tying it to partners' land, power and equipment.
Most of the cost blocking a switch to a different inference accelerator comes not from contracts but from inertia. Accumulated CUDA kernels and optimization know-how, the operations team's learned expertise, and integration with existing pipelines form a wall far higher than the actual invoice, and the cost-performance gap Google presented is the first time that wall's height has been turned into a number that can actually be compared -- meeting the conditions of the switching-cost lens exactly.
This news is often read as a chip-performance showdown -- 'Google challenges Nvidia.' But from a buyer's standpoint, what matters isn't which chip is faster but what your own organization's switching cost is, and without knowing that number you can't get the fact that an alternative exists reflected in price at the negotiating table. An alternative doesn't have value only once you actually switch to it -- it has value the moment you can prove that you could.
Decision prompt — Pick one or two of your most expensive inference workloads, move them above a framework abstraction layer, and within 90 days build a 'switching cost statement' by measuring identical-conditions benchmarks and migration effort on both TPU and your existing GPUs. Use that number as your position paper for the next GPU volume negotiation, and lock a rule into your architecture standard that new code must not call vendor-specific kernels directly.
About the framework
When the cost of switching to another product is high, customers stay put even when somewhat dissatisfied with what they're using, because invisible costs -- data migration, retraining, integration with existing systems -- are often much larger than the sticker price. Switching cost thus becomes a moat that keeps customers locked in for the incumbent, and a wall that's hard to scale for new entrants.
SemiAnalysis
Anthropic details distillation campaigns from Alibaba, Moonshot AI, and DeepSeek
Anthropic disclosed that its models are being distilled at industrial scale by Chinese companies including Alibaba, Moonshot and DeepSeek, and US government agencies have officially raised concerns about six Chinese AI firms replicating frontier models. Meanwhile, Y Combinator's Garry Tan has argued that the American open-weight camp should also distill frontier models for its own use -- the same technique being discussed simultaneously as an attack and as an industrial-policy tool.
Why a particular company stays ahead of others in the same industry has been explained by an internal resource -- model weights -- but distillation directly demolishes that resource's core condition: inimitability. If you can pull outputs through an API, a large share of the capability gets replicated without ever stealing the weights, so the premise that the resource must be scarce and hard to copy erodes the very moment the service is made public.
This news is often read as a geopolitics story about Chinese technology theft. But the resource-based view reaches a far more uncomfortable conclusion: even blocking it with export controls or lawsuits doesn't change the fact that model capability itself has already become a short-half-life asset. If that's true, moving the moat from the model to assets that can't be replicated with a single API call -- robot training data, exclusive contracted data, power and distribution channels -- is more urgent than playing defense.
Decision prompt — Rewrite this quarter any business plan that assumes the model itself is the moat, apply output watermarking, anomalous-call detection and contractual anti-distillation clauses immediately to high-value model APIs, and, within the same budget, reallocate your resource portfolio toward securing exclusive data and distribution-channel contracts. For assets that only come from the physical world, like robot training data or industrial field data, locking them up through partnerships now is cheaper than the price you'll pay a year from now.
About the framework
This view locates the reason a company keeps outperforming others in the same industry inside the company rather than in its market position -- the answer being resources and capabilities that others don't have. Not just any resource qualifies, though: it must be valuable, rare, inimitable and non-substitutable (VRIN), which covers things you can't simply buy -- patents, proprietary data, exclusive algorithms, talent.
TechCrunch
Based on 103 items over 7 days