Good morning everyone,
Building on last week’s geopolitics-and-markets lens, this edition tracks four shifts: AI’s industrial build-out, distribution moving into AI assistants, and what an AI “bubble pop” could look like. Plus others relevant news.
The AI build-out goes industrial
AI has broken out of the software world and is now an industrial project. Think power plants, grid connections, mega data centers, specialist chips, and long-haul logistics all scaling together. The practical constraint isn’t clever code—it’s electricity, sites, and supply chains. Capital is shifting toward regulated grid assets and long-dated offtakes; private infra funds are treating data-center platforms like “utilities with upside.” The near-term watch-outs are simple: can we connect to the grid fast enough, and will promised power actually arrive on time and on budget?
Mega-campuses are becoming standard. OpenAI, Oracle and SoftBank release new multi-billion-dollar plan to add 5 new data center to Stargate, raising the planed capacity of the site to 7 GW (a gigawatt powers ~750k homes) to run future models—locking in demand for land, power, and cooling for years.
Electricity is the bottleneck. A Bloomberg study shows that data-center demand is already nudging wholesale power prices up for households and businesses, forcing policy choices on who pays and how fast capacity expands.
Nuclear is back in the conversation. Forecasts now include a nuclear build-out of $350B to supply steady “baseload” power; “SMRs” (small modular reactors) are being discussed for on-site use, though approvals and costs are challenging.
More suppliers for compute capacity. Cloud giants (often called “hyperscalers”) are diversifying supply: CoreWeave signed a long-term $14B deal with Meta to pre-reserve high-performance computing.
Why it matters: AI demand is now tied to power systems, real estate, and regulated assets.
Search is fading; assistants are the new storefront
Buying journeys are starting to move inside assistants. Instead of clicking results and filling carts, users will ask an agent to find, compare, and buy—often in one conversation. That requires two things from companies: (1) be legible to agents (structured product data, APIs), and (2) earn consent for using conversational context. The marketing shift is from SEO to “AIO” (Artificial intelligence optimization) and from web conversion to in-chat conversion.
Chat becomes point-of-sale. New open protocols let merchants plug in so you can complete instant checkout in the conversation (fewer clicks, fewer abandoned carts). This sidelines parts of search and marketplace discovery.
From SEO to “AIO.” Instead of optimizing for Google results, brands try to appear in assistant answers. That means using your conversation context - raising privacy and consent questions under GDPR/CCPA.
Ads inside the assistant. Meta plan to personalize content and ads using chat context (what you asked, when, and why). A move that could be follow by OpenAI which seems to be staffing up to turn ChatGPT into an ads platform.
Implications: Optimize for assistant placement (like premium shelf space), build consent/audit trails, and earmark performance budgets for assistant placements with in-chat conversion metrics.
Quick hits (skim-friendly)
Agents in production: Citi pilots task-completing agents for 5,000 users; Microsoft allows personal Copilot at work with IT controls (=you can now use Copilot without IT having to pay for it); Stellantis partners with Mistral to integrate AI in supply chain.
Model race: Anthropic’s releases new Claude Sonnet 4.5 showing great new capabilities; OpenAI’s Sora 2 (video + social creation) - a hyper realistic AI generated Tiktok which raises concern on disinformation.
Robotics & energy: DeepMind’s Gemini Robotics 1.5 helps robots plan multi-step tasks; energy traders get 7-month hourly demand forecasts.
Mega-valuations: OpenAI secondary at $500B; Black Forest Labs (German Image Generation) targets $4B; DeepL (German Translation Company) explores US IPO; xAI (Musk) at $200B.
Society & governance: Stanford warns of “workslop” (low-quality AI created workshops) hurting trust/productivity; OpenAI adds teen parental controls; high-profile AI-and-apocalypse narratives still shape debate, especially from Peter Thiel - a close ally of POTUS.
Is (or when) the Bubble gonna pop?
The last couple week have been market by more and more concerns on the AI Bubble. Analyst point to classic bubble markers - record investment flows, lofty earnings expectations, heavy capex tied to power/regulation. A “pop” wouldn’t mean AI goes away; it would mean valuation compression while the infrastructure and use cases continue maturing. The likely trigger sequence: revenue misses → capex cuts → grid/permitting delays → tighter financing, with share prices resetting before the next up-cycle.
Record capital inflows. About $193B of venture capital went into AI in 2025; the biggest tech firms are committing trillions to data centers, chips, and models - even though today’s incremental revenue from AI is still modest.
High expectations. Markets assume smooth, fast monetization; JP Morgan AM worries earnings misses could trigger broader re-rating.
Macro linkages. AI capex depends on regulation and power; pullbacks could hit cloud providers and app layers first and ripple into indices and GDP.
Infrastructure constraints. Limited power, long build timelines, skilled labor shortages; some banks warn nuclear optimism may be ahead of itself (gas/renewables likely bridge this decade).
Why it matters: Prepare for valuation volatility around earnings/capex and for scenarios where power or permits slip against plan; stress-test exposures to hyperscalers, semis, and AI-heavy clients.