iiinova — where your digital self keeps evolving

Why do I even bother with Gemini?

000

7/7 Dots is a management layer over the agent economy. One point of contact that spawns the work, voice included, built on enterprise scaffolding the startups don't have. That's the story the name debate is hiding. Your move, Anthropic. Your move, Facebook.

000

6/7 The question is economics. Always-on agents on Astra, their most expensive model, each with a cloud computer and browser. First dot free, and chats don't count against usage limits. Subsidized at launch. If the enterprise wedge works, the consumer side is the flywheel. Think Iphone Adoption

000

5/7 The Instinct comparison misfires. Instinct is a personal-assistant startup racing for your phone. OpenAI already has enterprise scaffolding: specialist dots with org-level credentials, Microsoft Agent 365 for governance, internal runs in procurement, invoicing, support.

000

4/7 The real unlock is the GPT Live voice model underneath. Duplex voice makes calling a dot feel like a phone call with a colleague. Start jobs from a car, check in with a ten second call. Natural interaction plus work from anywhere is what makes always-on agents actually usable.

000

3/7 I've been testing it at work and it feels like a very competent engineering manager. Instead of prompting junior engineers one at a time, you have a single point of contact that spawns multiple Codex sessions and tracks them to done. That's OpenAI moving up the stack.

000

2/7 In one breath, Dots is a fleet of always-on agents. Each one runs 24/7 on its own cloud computer and browser, learns your preferences over time, and works toward your goals. Reach it from ChatGPT, Slack, or Teams. 4,000+ app plugins. It can drive your own laptop with permission.

000

1/7 OpenAI spent most of its Dev Day on a product that most coverage is still getting wrong. GPT-6.1 Sol got a fraction of the stage, Dots got the rest. The name and the promo video are the least interesting parts. This thread is why the product matters more than the model.

000

Zoom out and the position is clear. The company that spent 16 years teaching the web who is a bot is now the only seat that can decide which AI agents get in, and charge them for getting in. Gate and toll, same building. That is the quietest major move in the agent economy.

000

7/8 The risk is neutrality. Cloudflare insists the rails are open standards and that customers pick their own identity and payment providers. But the owner of the gate has an incentive to set the fee schedule. If labs route around or the detectors over-block, the bet reverses.

000

6/8 And the toll is on the road already. Agents prove who they are with Web Bot Auth. They pay with Wallets on the x402 standard, fractions of a cent per fetch. A site that never made money on ads can charge per page view and break even. The meter reads a trillion requests a day.

020

5/8 The sandbox is the other half. Workers runs code in V8 isolates across 330+ cities. Their agent-only browser, Kitesurf, spins up on Workers per request and gets discarded. Every agent on the web can run in a Cloudflare sandbox, on a Cloudflare meter. That is the gate.

001

4/8 Cloudflare's own framing is blunt. An agent doesn't render your CSS or click your ads, but there is a paying human on the other end. Block the agent and you block the customer. That one line is the whole business case, and it explains everything they have shipped since August.

000

3/8 What people are missing: AI agents look exactly like first generation spam bots. No CSS rendered, no ads clicked, thousands of fetches in minutes, the same fingerprints as the Russian botnets they spent a decade learning to stop. Their bot score already flags most of them.

000

2/8 The seat: Cloudflare analyzes more than 1 trillion requests a day, sits in front of 20% of the web, and is the first line of defense for a lot of storefront traffic. In May 2026, automated traffic overtook human traffic on their network. Their forecast had been late 2027.

000

1/8 Cloudflare has spent 16 years selling the same thing: a toll booth at the edge of the web. DDoS came first, then bots came, and now AI agents are the traffic. Their own 2026 numbers explain why this seat matters more than any model launch this year.

000

At ~$30B you are not buying $140M of half-year revenue. You are buying the 2028-29 steady state, roughly $8B a year, if every delivery lands. Watch the Anthropic financing, Monarch, and anchor prepayments. Underwrite the credit, not the narrative.

000

What is real: real MW in low-cost power regions locked before the AI land rush, the Microsoft relationship, and $137M of balance sheet debt against $6.5B of customer prepayments. Everything else is a 2028 binary: Monarch 2GW on schedule, or the book starts to slip.

000

Jensen Huang said in September that Nscale would have ~300k GPUs online by end 2026. A year later, with the Microsoft and Anthropic deals signed, the filing reports 25k active. The number selling the IPO is the supplier's, not the registrant's.

000

The candid bits are rare in an S-1. 2024 revenue came from hosting crypto miners. The biggest 2025 customer, 73% of the year, is not named. Auditors flagged going concern doubt, resolved by a plan to defer or cancel capex. Same document sells 'exceptional delivery certainty'.

000

The moat they pitch is power: ~70% cheaper than major US markets, owned sites in Norway, Portugal, Iceland and West Virginia. Fair. But GPUs are the big cost line at the same (rising) price as everyone else, and the revenue live today sits in colocation shells in Sines.

000

So ~85% of the backlog is two counterparties, one of which is a direct competitor in Azure. The rest includes Figure, a robotics company with ~$19B raised committing $3.5B for up to 100k GPUs from late 2027. This is a credit book being sold as a revenue book.

000

The book leans on two anchors. A $44.6B Anthropic deal, signed three weeks before the filing, runs off a West Virginia gas microgrid campus whose first 2GW is not online before H1 2028. The filing admits no binding financing commitments. Microsoft is the other ~42%.

000

George Z Lin @gzlin.bsky.social · 16/09/2026

Nscale filed for a NYSE IPO with a $103B contract backlog and $140M of half-year revenue. Only $2.6B of that backlog is active. The rest is campus that does not exist yet and financing that is not committed yet. The filing is worth a read.

000

DeepMind paper Dream-RSI converts logged discovery traces into frozen replay simulators enabling cheap off-policy dreaming to efficiently iteratively evolve executable exploration policies, foundational step towards self-improvements

010

9/9 The lesson is that span of control runs through the human over people, the human over agents, and the agent over agents. The winners keep every span small and every handoff clean, not one flat chart with a magic middle.

100

8/9 The people who actually run large agent stacks do not let one person or one coordinator watch everything. They group the work into small clusters with clean handoffs and pass only what the next level needs.

000

7/9 And its span is tighter than a human's. People compress context with trust, norms, and memory. An agent has none of those shortcuts, so every fact has to live inside a finite, expensive window.

000

6/9 A lead agent can only hold so many subagents in context before it stops telling them apart. Past that it misroutes work, re-reads things, and loses its place in the plan.

000

5/9 The layer in the middle is an agentic system. A lead agent plus a stack of subagents. It hits the same wall the human manager hits, except this wall is written in hardware.

000

4/9 Agents do not flag problems the way people do. They proceed, and when they fail they fail silently or confidently wrong. So the human carries the full detection load for each one, and that load does not scale.

000

3/9 The flattening fantasy has a junction it skips: the human on top. One person can only directly manage a handful of agents before oversight collapses.

000

2/9 But span of control was never a headcount. It is how much context one node can hold while still making good calls on every node below it. That is why the human range sits around five to fifteen.

001

1/9 The hottest org move is flattening. Jack Dorsey wants 6,000 people reporting to him with AI carrying the context. Jensen Huang runs 60 direct reports. Everyone is pushing span of control further.

000

So the 'AI gateway' was three assets under one name, and the market priced them separately against three buyers. The durable value sits at the meter and at agent identity, now held by incumbents not a venture pure player. The open question: does Stripe keep the router neutral?

000

The knock-on is the neutral middle giving way. Observability got liquidated fast: Langfuse to ClickHouse, Helicone to Mintlify, Galileo to Cisco, then Arize. What was a venture opportunity is now mostly LiteLLM plus the platform and hyperscaler gateways.

000

None of the buyers paid for the software profit and loss. Each paid a strategic price for the wedge that slice gives into the agent economy. Hence the multiples do not line up: near 70x at the meter, low 20x at observability, single digits on a small base at the checkpoint.

000

Slice three is the trust layer. Dynatrace is buying Arize for $915M, the largest move in AI observability, betting the category clears $10B by 2030. The point is to reach the engineer before an app ships and to own the trace of what a model decided, because the trace is the loop.

000

Slice two is the checkpoint. Palo Alto closed on Portkey in May, reportedly around $120 to 140M, and rebranded it the Prisma AIRS AI Gateway. The tell is in its own traffic data: MCP traffic on its firewalls ran from 11 percent to 41 percent of volume in about six months.

000

Slice one is the settlement layer. Stripe is taking OpenRouter for over $7B, about 70x trailing revenue, a 5x jump off the $1.3B round from May. The thesis is that metered pricing is the native business model of the AI era, and Stripe wants to own the point where the meter sits.

000

The 'AI gateway' is being quietly retired as a category. Over the last eight months it split into three separate businesses, each bought by a different incumbent. Payments took one slice, security another, observability a third. Here is the map.

000

8/8 Open source security was built on keeping bugs quiet long enough to fix them. Agents ended that trade. The control now is throughput: continuous releases, mitigations that propagate in minutes, a trusted channel so the right people hear the hint first. Throughput is the patch.

000

7/8 Two answers fit this physics. Ship continuously in public: Chrome ships two security releases a week, the kernel defers fixes at most seven days. And put protocol layer mitigations live while the real fix lands. Cloudflare did that for Log4shell in 2021.

000

6/8 Fixing it in secret plugs the wrong leak. GitHub's temporary private forks cut off CI, hold one PR, and enroll reviewers one at a time. And the patch being secret matters less than the bug description not reaching an attacker's agent. OSS has no trustworthy channel for that yet.

000

5/8 A 2026 paper calls the asymmetry bugonomics. Exploit generation scales with cheap compute. Maintainer validation, triage, and release throughput stay flat. Disclosure volume that used to arrive over a decade is arriving in a month, much of it machine generated.

000

4/8 Mean time to exploit is now negative. Exploits ship before patches do. Around 2018 the number was near 63 days, and it crossed zero in 2024. One 2026 CVE went from public advisory to first exploitation attempt in nine hours, with zero public proof of concept anywhere.

000

3/8 The research backs it up. A GPT-4 agent given a CVE description exploited 87 percent of a 15 vulnerability benchmark. Given nothing, 7 percent. Embargoes existed because the description was the crown jewel. The description is now the attack surface.

000

2/8 You can point a model at the affected code and it writes a working exploit against a live endpoint, on a laptop with a prompt, in under a minute. The same agents that find the bug weaponize it in the same session. Finding, fixing, and exploiting used to be three activities separated by months.

100