본문 바로가기
IT Strategy Weekly

Oct 18 – Oct 24, 2026

Korean original

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

This week’s real shift was not the race for model performance. Value moved to the question of who controls the floor that AI runs on. Amazon is pulling PyTorch onto its own chips, Huawei and DeepSeek are completing a Chinese vertical stack, and Anthropic is struggling to secure compute. Money and bargaining power are being reallocated below the model, not above it.

The strategist's view

Next year's AI profit and loss will depend less on the time spent choosing a model than on the time spent making models swappable.

This week Replit switched its default model from Anthropic to OpenAI, and it took a one-line announcement. Anthropic, by contrast, holds huge funds but cannot get enough compute from cloud providers, and its partnerships have become tangled. Amazon is drawing developers onto its own chips with PyTorch-compatible tools, and DeepSeek is laying a software stack on Huawei chips to build a Chinese vertical ecosystem. Taken one by one, these are four small stories. Taken together, they point in one direction: the model on the upper layer is becoming a part that can be swapped quickly, while the chips and serving infrastructure on the lower layer have become scarce resources you have to queue for.

That means the question many companies have asked for the past two years, 'AI strategy = which model should we use?', has absorbed the most executive time for the decision that loses its value fastest. A company that needs six months to switch models and one that needs two weeks have completely different bargaining power with suppliers, even if they use equally capable models. When a price rise arrives, the first has no choice but to accept it. The second can simply move elsewhere.

So the question for next year's AI budget meeting should not be 'which model is smartest?' but 'how many days does it take to change one line of model code in our system?' An organisation that does not know this number has no strategy. One where the number exceeds a month is already locked in. At the model layer, spend only on abstraction and redundancy. Concentrate your people on serving, evaluation and agent permission control, where no one has yet set a standard. That is the only place capabilities that cannot be copied will remain.

Cases through a framework 3

Value Migration

Amazon is taking its biggest shot at Nvidia yet with new TorchNeuron dev tool

Amazon released TorchNeuron, a developer tool that lets code written in PyTorch run almost unchanged on its Trainium chips. Separately, it overhauled Trainium's training software stack to match the native PyTorch experience. In the same week, DeepSeek announced support for a software stack (TileLang, DeepGEMM and others) for Huawei's Ascend chips, completing a vertical ecosystem for the Chinese domestic market. Meanwhile, despite large investments and IPO preparations, Anthropic has not secured enough compute from cloud providers, and its partnerships are becoming tangled.

The AI industry is at a point where the source of profit is actually moving. The performance gap between models is narrowing, while the chips, power and serving stacks needed to run them are in short supply, so that even a top model company such as Anthropic is reduced to begging for compute. Amazon is trying to pry developer experience away from Nvidia's CUDA, the single gateway it has been tied to, by offering a PyTorch-compatible detour. These are the classic conditions in which revenue figures and the real flow of value diverge.

This news is usually read as a chip-competition story: 'Amazon finally challenges Nvidia'. But a market-share reading cannot explain why Amazon started with developer tools rather than hardware, or why DeepSeek, which is not a chip company, is releasing kernel libraries. Through the lens of value migration, both moves are attempts to occupy the chokepoint that decides who collects the tolls now being charged below the model layer. Companies that chase only model performance will be hollowed out, with revenue intact but margins draining away.

Decision prompt — When you draw up next year's AI budget, separate model licences from compute instead of treating them as one block. Within the next quarter, verify with real benchmarks whether at least 20 to 30 percent of your inference workloads can run on alternative accelerators such as Trainium or Inferentia. At the same time, have your legal team check immediately whether your current model supplier contracts hide an 'SLA exemption if compute procurement fails' clause. If they do not include capacity guarantees, add them at renewal.

About the framework

Industry profits do not stay in one place. They flow towards business models that better meet what customers really want, because when technology or customer needs change, the places that used to make money empty out and value moves somewhere else entirely. So a company whose revenue still looks healthy may already be losing value and slowly hollowing out. Winning depends not on where you make money now but on reading where profit is flowing.

TechCrunch

Wardley Mapping

Designing an AI serving harness built to last

An analysis argues that competitive advantage in enterprise AI is shifting from the performance of the model itself to the serving harness architecture that deploys models in production reliably and cheaply. In the same week, Microsoft warned customers to treat agents built on frontier models as a security risk on a par with insider threats. Two other debates emerged alongside: 'context-aware AI', which holds that general-purpose models alone cannot capture business constraints and operational context, and a counter-argument that AI cannot replace engineering teams.

This is a moment when the components of the AI value chain can be ranked by maturity, and the immediate issue is what to build and what to buy. Foundation models are already close to interchangeable commodities, so Replit could swap its default model from Anthropic to OpenAI with a one-line announcement. Serving harnesses, evaluation and governance, by contrast, are still at the custom stage that each company has to build by hand. On the map, the line between where to invest in-house and what to buy like a utility is sharp.

This news is usually read as practical advice: 'AI operations are hard, so engineering needs to get better'. But that reading leads to the waste of putting equal effort into every layer. Seen on the map, the judgement becomes simple in the opposite way. Do not spend on the model layer, which is already commoditised; just secure the ability to swap. Concentrating your own people on the serving, evaluation and agent-control layers, where no one has yet set a standard, is the only investment that cannot be copied.

Decision prompt — This quarter, produce a one-page map that splits your AI stack into four layers (model, serving, evaluation and governance) and classifies each as 'buy' or 'build'. Wrap the model layer in an abstraction layer that can plug in two or more suppliers, and cut switching cost to two person-weeks or less. Conversely, do not outsource the serving harness or agent permission control. Assign a dedicated internal team to them, and apply the same access logs and permission reviews to agent accounts as to human employees.

About the framework

To set strategy you first need to see the landscape. Wardley Mapping places customer value at the top and links the components it needs in a chain below, positioning each along a horizontal axis by maturity (genesis → custom → product → utility). This makes visible on the map what will soon be commoditised and fall in price, and where to invest in-house versus what to hand to outside providers, so you can move on the basis of the terrain rather than gut feeling.

MIT Technology Review

Ambidextrous Organization

Meta wants to pay for its broad AI buildout with a newly created class of stock

Meta is pursuing a plan to issue a new class of shares to fund large-scale AI investment. In the same week, Tesla's third-quarter net profit fell 43 percent on aggressive capital spending and soaring costs, OpenAI prepared to raise $30 billion at a $1.25 trillion valuation, and Blackstone said concerns about AI infrastructure debt were overblown.

This is a moment when an existing business that makes money today and an exploratory business that does not yet turn a profit collide head-on within one set of financial statements. Meta's push for a new class of shares is an attempt to escape a structure in which its AI capital spending is judged by the profit metrics of its advertising business. Tesla's 43 percent drop in net profit shows the reverse: what happens when the two activities are mixed in one income statement with no separating mechanism.

This news is usually read as a financing story: 'how will Meta fund its AI spending?' But a financing reading misses why it needs a new share class rather than ordinary shares or corporate bonds. Through the lens of the ambidextrous organisation, this is a question of organisational design, not financial technique. It reads as a mechanism to achieve separation and integration at once: shielding the exploratory business from the evaluation logic and investor expectations of the existing business, while keeping both bound together at the executive level.

Decision prompt — Restructure so that AI investment is not folded into existing business units' profit and loss. Report it separately each quarter, with its own investment account and its own metrics (learning speed, unit inference cost and internal adoption rate, not revenue). Place approval authority with a committee reporting directly to the CEO, not with business unit heads. At the same time, give that exploratory budget an explicit expiry: 'transfer to the core business or shut down within two years'. This pre-empts both failure modes, an innovation unit that drifts on its own or one absorbed by the existing business.

About the framework

Lasting companies must be able both to run the existing business that makes money today efficiently and to explore new businesses that open up the future. The two demand opposite cultures, metrics and time horizons, so they constantly clash within one organisation. The key is to achieve separation and integration at once: shield the exploratory unit from the logic of the existing business, while linking the two at the executive level.

TechCrunch

What to watch

Based on 121 items over 7 days

Past issues