본문 바로가기
IT Strategy Weekly

Oct 4 – Oct 10, 2026

Korean original

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

The real shift this week is not the race over model performance. AI’s value has moved one layer down, from ‘who has the smarter model’ to ‘who controls the protocols through which agents trust and connect to one another, and the permission and power foundations beneath them’. And the very property that makes value grow with connection is also becoming the biggest risk.

The strategist's view

The decisive ground in AI competition has moved from model performance down to 'connection rights', and companies that do not design those rights now will live as tenants inside someone else's protocol two years from now.

Read this week's articles as a ranking of model performance and you see nothing. In the same week, an attack emerged that spreads along trust in an agent-to-agent communication protocol, Apple blocked agents' disk access at the OS level, and a report found that most enterprise agent projects run aground because they cannot reach organizational knowledge. Together, this means the competitive bottleneck has moved from 'what can it generate?' to 'what is it authorized to access?' A study finding that AI coding agents pour out code but stall at human review, so actual software output does not rise, says the same thing from another angle: output volume is no longer the scarce resource. Verification and permissions are.

In protocols whose value grows with connection, whoever gathers mass first wins, so Apple, Amazon and agent startups are all pushing for the standard position now. The problem for adopting companies is that joining after this race ends will be too late, while tying everything hastily to one protocol now will leave them unable to get out later because of the cost of reconnecting tools and redesigning permissions. The answer lies in between. First build, as your own asset, the permission layer and adapters that control what agents can reach, and design it so any protocol can be swapped in on top.

The same logic repeats in power. News that batteries have become cheaper than gas turbines looks like environmental news, but in reality it lifts a constraint on where and how fast AI infrastructure can be built. Connection rights above, power modules below: this year, these two must take the place model names used to occupy in AI strategy documents.

Cases through a framework 3

Network Effect

MCP for agent-to-agent comms may be the riskiest protocol you've never heard of

A new prompt-injection attack was identified in the MCP protocol, which AI agents use to call one another and share tools on internal networks. Malicious instructions spread not through the model itself but along the trust relationships between agents. In the same week, Nous Research was valued at $1.5 billion for its business agent products, and an analysis found that two-thirds of enterprise agent projects fail to reach commercial use because they cannot connect to organizational knowledge. Apple reworked macOS's Full Disk Access permission system to stop AI agents from accessing messages without authorization.

MCP is a classic protocol race: as more agents and tools connect, each agent can do more, so whoever joins first gains more. Once woven into internal workflows, it is hard to switch because of the cost of reconnecting tools and redesigning permissions. And no protocol has yet clearly crossed critical mass, while Apple, Amazon and agent startups are all going for the standard position at once. The conditions for applying the framework fit exactly.

This news is usually read as 'another security vulnerability has been found'. Through the lens of network effects, however, it is not a bug report but news about the state of the protocol race. In a structure where connection creates value, the path along which trust spreads and the path along which malicious instructions spread are physically the same. So the protocol that spreads fastest is also the one with the widest attack surface, and in the end, whoever shows 'how to connect safely' takes the standard.

Decision prompt — Do not expand internal agent connections across the board yet. This quarter, first build a connection map and call logs showing which agents access which tools and data, then enforce the same least-privilege and identity-verification rules for agent-to-agent calls that you apply to people. And instead of betting everything on MCP alone, add an adapter layer so that you do not have to rebuild workflows wholesale if the protocol winner changes.

About the framework

Some technologies get better for existing users as more people use them, because, like communication protocols or platforms, their value grows as they connect with others. So the leader pulls further ahead and the market tips toward a few winners. They win not because their quality is slightly better but because people have already gathered there.

Ars Technica

Platform Envelopment

Apple is reportedly partnering with LG to launch a smart lock, thermostat, and doorbell

Apple was reported to be preparing a lineup of smart locks, thermostats and doorbells with LG. Analyses the same week noted that Apple is shifting its weight toward partnerships and interoperability standards instead of holding to its old fully closed loop, and that in the process it is pressuring competitors such as Amazon to rework their agent strategies.

The smart home buyers Apple is targeting overlap almost entirely with people who already use iPhones. It can reuse existing infrastructure such as identity, payments, notifications and on-device AI as is. And a bundle that lets users handle locks and doorbells together from the iPhone lock screen feels entirely natural to them. All three conditions for enveloping an adjacent market are met.

This news is usually read as a product launch story: 'Apple is finally making its own smart home hardware'. Through the lens of envelopment, though, Apple is not trying to be number one in doorbells. It is trying to pre-empt the touchpoint through which agents in the home speak. Individual IoT makers are pushed out not by losing on product quality but by losing pricing power as their products are repositioned as one slot in someone else's bundle.

Decision prompt — If you sell smart home or IoT hardware, do not frame 'should we join the Apple ecosystem?' as an either-or question. Account separately for revenue from the devices themselves and revenue from the data and services those devices generate, and first find out which one is profitable. Then make it a red line in partnership negotiations that, even if you are folded into a bundle, the customer relationship and sensor data stay in your own app.

About the framework

A platform swallows an adjacent platform's market not by building a better product head-on but by bundling adjacent features onto its existing user base and enveloping the whole market, because overlapping users and shared infrastructure let it enter a new market at far lower cost. So even without being the best in each market, the ability to bundle adjacent markets decides the contest.

TechCrunch

Disruptive Innovation

Batteries are now cheaper than natural gas turbines used at many data centers

The surge in data-center demand has actually driven battery prices down, making battery-based power configurations cheaper than the natural gas turbines used at many data centers. The same week, the IEA noted that the world has entered a phase in which security, affordability and cleanliness are achieved together through electrification, and there was also news of semi-solid-state batteries entering commercial production and the development of next-generation power sources such as high-temperature gas-cooled reactors.

Batteries still trail gas turbines on traditional generation metrics, long continuous output and large-scale baseload power, but they lead on other axes: installation speed, modular expansion, siting constraints and carbon cost. Incumbent power suppliers are optimized for large, long-term contract customers and have little incentive to enter a corner market sold in small pieces. And batteries' price and density are improving faster than the reliability data centers demand is rising. The three conditions for disruption fit.

This news is usually read as an energy story: 'the green transition has finally become economical'. Through the lens of disruptive innovation, however, this is not an environmental story but an event that changes the constraints on data-center siting and expansion speed wholesale. The moment power can be added in modules, the premise of 'building only where the transmission grid reaches' falls away, and the terrain of AI infrastructure competition itself is redrawn.

Decision prompt — Recalculate your next three years of compute expansion plans together with how you will procure power. Secure renegotiation or shortening options on long-term power contracts set on a gas-turbine basis, and run a battery-centered pilot at one new expansion site to gather your own data on actual utilization and peak response.

About the framework

What topples incumbents is usually not better technology but worse technology: a cheap, simple, lower-performing product starts at the low end or in niches that mainstream customers ignore, then improves and moves up. Christensen's paradox is that incumbents lose not because they are incompetent but because they practiced rational management, faithfully serving the needs of profitable customers. That is why they abandoned the low-end corner, and why they are late.

TechCrunch

What to watch

Based on 123 items over 7 days

Past issues