• On About AI
  • Posts
  • Bubbles Build the Future: Bezos’ Optimism + This Week’s Agent Stack Shifts

Bubbles Build the Future: Bezos’ Optimism + This Week’s Agent Stack Shifts

Speculation funds infrastructure. Your job is to turn that runway into policy, telemetry, and compounding advantage.

Author

Joao Coutinho
October 10, 2025

Jeff Bezos’ Turin talk landed a simple paradox: AI is a bubble—and a revolution. That’s not a contradiction; it’s a blueprint. Bubbles pre-pay the infrastructure everyone else later takes for granted—data centres, energy coupling, agent frameworks, and evaluation pipelines. Our edge isn’t guessing winners; it’s owning where AI runs, how it’s governed, and how it’s observed. Full Italian Tech Week replay

Worth the watch, he has several interesting points. I’ve curated the ones that sparked my curiosity the most. You know, those that make you dream about the future.
The first one is about Entrepreneurial Optimism, “Optimism almost to the point of delusion”. I find this one very powerful, to beat the fear of the unknown we need to be extremely optimistic, this is crucial for overcoming challenges and driving innovation. The second quote is about Reality Wins, “Reality is undefeated. Reality wins every time.” Bezos emphasises pragmatism and self-awareness, urging entrepreneurs to build on actual customer needs rather than imagined ones. If you focus on real needs, you’ll have bigger chances of success. The third one, and probably my favourite, is about Space and the Future. I think this triggers our Sci-Fi senses and Star Trek vibes. He says “Space will make Earth better”. Bezos envisions gigawatt-scale data centres in space within 20 years, enabled by uninterrupted solar power and lunar resources. He predicts millions living in space due to rapid advances in robotics and engineering. I believe this really makes sense, cooling would be way more efficient and solar power would make this data centres even more efficient. Can you image commuting in space? instead of 2 days/week at the office you spend 3weeks in space. I could imagine future generations living permanently in space colonies. This is very futuristic, just enough to makes dream about it.

TL;DR

  • Bezos’ take: AI is a bubble and a revolution—good, because bubbles pre-pay infrastructure.

  • Platforms moved: Google launched Gemini Enterprise (the “front door” for AI at work); OpenAI shipped Apps SDK + AgentKit (zero-to-agent faster); Microsoft doubled down on governed agents.

  • Macro signals: Antitrust remedies will shape AI distribution; chip geopolitics continues to redirect compute capacity.

The Brief

Google rolls out Gemini Enterprise (front door for AI at work)

  • What happened: Google launched a unified enterprise AI subscription—agent building, Workspace tie-ins, admin controls. Pricing starts ~$30/seat/month (annual). Early customers named (e.g., Gap).

  • Why it matters: Budgets will migrate from ad hoc experiments to a platform line item. Identity, permissions, and distribution (surfacing AI where people already live) become the wedge. This Microsoft Copilot 365 a la Google.

  • Do now: Pilot a read-only spend insights agent (CSV in → “Top 10 vendors by MoM delta?”; “Contracts expiring in 90 days?”). Keep no PII and ban data export in the first sprint.

OpenAI ships Apps SDK + AgentKit (from ideas to agents faster)

  • What happened: New SDK to build Apps inside ChatGPT plus AgentKit for building/evaluating agents with first-party components (routing, tools, evals). Public partner apps announced.

  • Why it matters: Time-to-pilot collapses. Your bottleneck moves from glue code to governance & observability (logging, evals, approvals).

  • Do now: Stand up an Invoice Triage Agent (classify → extract totals/vendor → flag >€25k for second approver). Create a 50-doc test set and track F1 ≥ 0.90 for totals/vendor; show cost/run.

Google’s Gemini 2.5 Computer Use model (agents that click UIs)

  • What happened: API model for browser/mobile UI control using screenshots + actions; Google claims benchmark wins at lower latency; dev docs live.

  • Why it matters: This is RPA-meets-LLM without heavy, per-bot licensing. Great for “click-ops” tasks (weekly exports, report pulls) where APIs don’t exist.

  • Do now: Automate a weekly browser export from a legacy SaaS → checksum → SharePoint/Drive upload → Teams/Slack summary. Enforce domain + action allow-lists and persist a step ledger (timestamp, action, selector, result).

Microsoft doubles down on governed multi-agent systems (Azure AI Foundry)

  • What happened: Microsoft introduced Microsoft Agent Framework and previewed Task Adherence (keep agents on-task) and prompt shields with “spotlighting” (detect prompt injection / risky behavior).

  • Why it matters: Governance is getting productized: policy-as-guardrail, audit-ready logs, and content safety inside the platform—not a bolt-on.

  • Do now: Add Task Adherence checks; implement a Spotlighting-style shield; prove it with a blocked off-scope tool call that returns a friendly error + remediation hint.

Antitrust remedies start shaping AI distribution power

  • What happened: Following the U.S. search case, DOJ secured “significant remedies” against Google’s monopoly practices; policy debate now includes AI bundling (e.g., pairing Gemini across surfaces).

  • Why it matters: Defaults and distribution—not raw model IQ—decide adoption curves. Expect contracts and platform UX to change as remedies bite.

  • Do now: Attach a No-Lock AI Addendum to RFPs: portability (export prompts, logs, evals), neutral IDP, BYOK, and a 30-day exit SLA with assisted migration.

Chip geopolitics: UAE green-lights vs. China crackdowns

  • What happened: U.S. authorisation advances large NVIDIA shipments to the UAE (reports: up to 500k advanced chips/year under a bilateral framework); meanwhile China tightens customs checks on NVIDIA-class imports.

  • Why it matters: Regional capacity and latency will shift as the Gulf adds supply and China enforces stricter gates. Your resilience depends on multi-region options and sanctions-aware procurement.

  • Do now: Map where your inference runs today; prepare a region failover plan; add a sanctions audit trail to vendor onboarding.

Bezos’ optimism works only if we pair it with distribution discipline and governed execution. Google’s front door (Gemini Enterprise) gives you reach. OpenAI’s AgentKit collapses build time. Microsoft’s guardrails keep you audit-ready. Antitrust will reshape defaults; geopolitics will reshape where your capacity lives. The winning posture is simple: pick your venues, instrument everything, and never lock yourself in.

Deep Dive

Apply AI: Europe’s Distribution + Sovereignty Play

Bezos’ Turin message—AI is a bubble and a revolution—perfectly aligns with news this week because Google is doing the two things that convert hype into durable advantage:

  1. building a distribution “front door” for agents at work (Gemini Enterprise + partner ecosystem), and

  2. laying concrete and cable in Belgium to power it (extra €5B into DCs + clean-energy PPAs).

  3. The EU’s new Apply AI Strategy is the policy that will glue all this together and enabling budget spending.

Distribution: the partner-powered “front door”

Google’s Gemini Enterprise reframes AI as an entry point employees already trust (Workspace, browser, mobile) with a marketplace of partner-built agents and integrations. In plain terms: the more “doors” into everyday work, the less the model brand matters and the more distribution wins. Google’s new partner note makes that explicit: solutions are meant to slot into Gemini Enterprise and go to market through that surface.

So what?

  • If you’re a vendor, you race to be the default agent for a job-to-be-done inside Gemini.

  • If you’re an enterprise, you’ll buy capability bundles (front door + agents + admin controls) rather than a dozen isolated tools.

  • If you’re in the EU, you also evaluate data pathways and identity (neutral IDP, exportability) before you enjoy the convenience of defaults.

Concrete: Belgium as a strategic AI zone

Google announced €5B of new investment in Belgium over the next two years—expanding the Saint-Ghislain campuses, adding ~300 full-time jobs, and signing fresh clean-energy agreements (Eneco, Luminus, Renner) including onshore wind that supports the grid. This is not a press release flourish; it’s siting strategy: cheap/cleaner power, political stability, and proximity to EU financial/regulatory hubs.

So what?

  • Latency & locality: more EU workloads can stay in-region without sacrificing responsiveness.

  • Sustainability optics: PPAs and heat-reuse potential strengthen ESG narratives for AI adoption in regulated sectors.

  • Sovereignty posture: Belgian/EU clients get a credible path to keep data + inference close, while tapping a global partner ecosystem. (Coverage outside Google echoes the €5B figure and strategic framing.)

Policy: Apply AI — Europe’s distribution + sovereignty play

The European Commission’s Apply AI initiative turns adoption into a roadmap: sector playbooks (health, energy, mobility, manufacturing, public admin), a stakeholder Apply AI Alliance to coordinate buyers and vendors, and an implementation spine (AI Office/Observatory) that pairs funding windows with practical guidance. Think of it as the how to deploy companion to the AI Act’s how to govern: make AI usable in-region, observable, and portable—on Europe’s terms.

So what?

  • Procurement acceleration: Sector guides + Alliance engagement shorten the gap from pilot to purchase in public services and regulated industries—expect clearer templates, KPIs, and proof requirements.

  • Sovereignty-by-default: Emphasis on EU residency, portability, and neutral identity means your prompts, logs, evals, and data flows should live in-region and export cleanly—reducing lock-in risk.

  • Vendor signal & funding alignment: Joining the Apply AI Alliance and pre-committing via the AI Pact signals readiness for EU buyers and can align your pilots with Horizon/Digital Europe calls—money and compliance moving in the same direction.

The Bezos connection

In Turin, Bezos argued that bubbles fund the physical and organisational scaffolding future products depend on—from more compute to new siting options (he even floated orbital data centres as a 10–20 year horizon where energy economics could flip). The point isn’t “space DCs tomorrow” but instead that distribution + energy are the real moats. This week’s Google moves are exactly that: one distribution surface (Gemini Enterprise) and one energy/infra bet (Belgium).

In Turin, Bezos argued that bubbles fund the physical and organisational scaffolding future products depend on—from more compute to new siting options (he even floated orbital data centres on a 10–20 year horizon). The point isn’t “space DCs tomorrow”; it’s that distribution, energy, and the rulebook are the real moats. This week lines up cleanly: a distribution surface (Gemini Enterprise), kilowatts underfoot (Belgium’s €5B buildout), and a policy rail bed (Apply AI) that makes adoption sovereign and portable. Translate the bubble into durable advantage—on your terms.

Next Steps

What to learn now?

That’s it for this week.

As AI moves from chat to real work, the questions we ask, and the discipline we bring to answering them, will define the next decade. This week’s signal is clear: choose a front door you can govern (Gemini Enterprise or not), anchor capacity in-region as Belgium scales, and adopt on Europe’s terms with Apply AI—portable, observable, sovereign. Prepare our kids, equip our organisations, and build with policy, telemetry, and an exit in mind. Clarity first. Action now.

Stay curious, stay informed, and keep pushing the conversation forward.

Until next week, thanks for reading, and let’s navigate this evolving AI landscape together.