VOL.2026.09.22 Today's call 4 min read

Two reports pushed compute costs up on the same day, and today's real openings sit in the middle layer: sensitive data, long-task state, and ad optimization.

Tuesday, September 22, 2026

—: The 2026 China AI Computing Power Development Assessment Report projects 87.9% growth in China's intelligent compute scale and says agents are reshaping demand structure — inference and agent workloads are becoming the cost driver.
—: Roughly $300 billion in AI-related financing surfaced, shifting compute buildout funding from internal cash flow toward capital markets and changing the risk structure of the infrastructure layer.
—: IAB flagged creators' significant role in LLMs while WPP and Fabulate warned that infrastructure and measurement have not caught up — the measurement gap is itself a position worth occupying.
—: Today's selected projects cluster into three middle layers: de-identification and restoration of sensitive data around models, state and context management for long-running agents, and agentic or pre-spend screening in ad ops and creative.

Today's Positive Direction

Compute is getting more expensive and more complex, and that pushes opportunity toward the layer that does not train models. Two market-context items point at the same thing: agent workloads are reshaping compute demand, while the money funding compute buildout is shifting from internal cash flow to capital markets. As the bottom layer gets heavier and pricier, the middle layer — how data safely enters and leaves a model, how a long-running agent remembers where it is, how repetitive work like ad ops and creative gets taken over — becomes today's densest band of opportunity.

Market Changes

3 items

The demand structure for compute has changed. The 2026 China AI Computing Power Development Assessment Report projects 87.9% growth in China's intelligent compute scale and states explicitly that agents are reshaping the demand structure of AI infrastructure. Judgment: cost pressure keeps moving from training toward inference and agent workloads, which raises the value of anything that cuts wasted calls, context rebuilding, or duplicate requests.
The money behind compute buildout changed hands. Roughly $300 billion in AI-related financing surfaced, with tech giants using external financing to move part of the funding and risk of compute buildout to Wall Street. Judgment: this does not immediately change application-layer opportunity, but it means infrastructure pricing and supply will be more exposed to capital markets — the application layer should not assume compute keeps getting cheaper.
Creators entered the models; measurement did not follow. IAB flagged creators' significant role in LLMs, while WPP and Fabulate warned that the related infrastructure and effectiveness measurement have not caught up. Judgment: the measurement gap is not noise, it is a clear position — whoever turns "what creator content does inside a model" into comparable numbers holds that link in the chain.

Projects Worth Expanding Today

4 picks

cover

Before sending sensitive material such as client lists, medical records or contracts to an external model, requests are routed through this proxy: real fields are replaced with realistic fake values, and after the model responds, the original values are restored locally. Users get usable model output while the real data never leaves.

It replaces manual de-identification — previously handled by legal, compliance or data engineering, slowly and with gaps. Judgment: this kind of pre-model middleware sits exactly where rising compute cost meets tightening compliance, and law firms, clinics and accounting practices are natural starting points. Integration method and restoration reliability still need verification.

awr

For long-running AI agents, it provides persistent work state and minimal context so agents do not rebuild the full context each turn.

It replaces developers writing their own state-management scripts. Judgment: this is the other side of the same coin as today's compute report — context rebuilding is both a cost and an error source, so abstracting it into a separate layer is a sound engineering direction. Integration method, where state is stored, and delivery form still need verification.

Adlyse

Paid media staff previously had to repeatedly adjust budgets, swap creatives and watch numbers across multiple ad platform back ends; Adlyse lets AI agents take over these daily operations and keep adjusting campaigns continuously. Users end up with a campaign state that has been executed and adjusted by the agent.

It replaces the daily manual optimization work of media buyers. Judgment: the notable part is not "AI does advertising" but that agents are taking over routine operations rather than only generating creative — which contrasts with today's IAB note on the measurement gap: the more agents execute, the more measurement matters. Which platforms it actually takes over still needs verification.

AI Creative Insights by Decode

Advertisers or agencies evaluating several creatives before a campaign open it, hand the candidate creatives to the system, and the AI predicts which is more likely to win, returning a pre-spend ranking or selection recommendation.

It replaces the "spend first, read data later" approach to creative testing, turning media budget into something that can be compared beforehand. Judgment: this is one end of the same trend as Adlyse — pre-spend screening on one side, in-flight execution on the other. The prediction basis, accuracy and human review step are not stated in public materials, and that is the first thing to press on.

Other Selected Projects

4 items

ainxt-os: an open-source framework released by India's National Payments Corporation; developers connect chosen models, tools and workflows to get a runtime foundation they can deploy themselves. Judgment: national payment and public infrastructure bodies are starting to build their own AI application foundation rather than only buying external model services — local deployment needs in regulated industries are worth watching.
AppGrowthKit: automatically generates App Store screenshot assets for different devices, languages and store versions. Judgment: a repetitive pre-release manual step, and charging per release version or asset set is a natural path, but it is a tool-shaped product that platform-native features could absorb.
blink: natural-language search for implementations and call relationships inside an existing codebase. Judgment: general-purpose code search is already covered by big-vendor tools; the opening is more likely in specific language ecosystems or compliant search over private codebases.
Flicka: a Chrome extension that generates publishable product demo assets in the browser, replacing screen recording, editing and voice-over. Judgment: indie developers shipping fast are natural users, but low delivery cost also means a short differentiation window.

Conclusion

No new model capability breakthrough appeared today; what appeared were shifts in cost structure and compliance structure. What is worth tracking is not "yet another AI tool" but the middle layer that takes on the dirty work between models and business: data de-identification and restoration, long-task state management, and pre-spend screening plus in-flight execution in advertising. Their common trait is that they do not depend on models getting stronger — only on models being used more.