We didn't come to BGC on a Thursday night to argue about servers.
Yet there we were โ thirty people on the second floor of a bar off 5th Avenue, San Miguel sweating in the Manila humidity, and a trader I'd last seen chasing SushiSwap APYs in 2020 was pitching something far stranger. He turned his laptop around and showed me a pricing page. Five dollars and seventy cents a month. One vCPU, one gigabyte of memory, seven days of maximum runtime.
"That's a pet," he said. "Not a service."
The table laughed. I didn't. Because the number wasn't the story. The story was that the machine economy had finally grown a runtime layer โ and nobody in that room, or in any of the vendor decks I read the following week, could agree on what shape it should take.
That disagreement is the most interesting macro signal I've seen since the ETF flows started printing in 2024.
Five roadmaps, five incompatible answers
Between August and September 2026, five players published five different answers to a single question: where should an autonomous agent actually live?
Google showed up with Cloud Run instances โ singleton, no autoscaling, a hard seven-day ceiling, and that $5.70 monthly tag for 1 vCPU and 1 GiB. It treats the agent as a pet: long-lived, stateful, always on.
AWS answered with AgentCore Runtime V2, now generally available. Each session runs inside a hardware-isolated microVM. It speaks MCP, A2A and plain HTTP. The cold-start P75 dropped from a brutal 5.4โ30 seconds in V1 to roughly 1.9โ2.0 seconds, and AWS added elastic memory reclamation so you pay for actual consumption rather than peak. Runtime Instances stretch to 14 days and can attach GPUs.
Cloudflare previewed @cloudflare/computer, where a Durable Object isolate runs the agent harness and containers are summoned only as tools. Both share a persistent filesystem. Isolates boot in under a millisecond across more than 300 points of presence, and Cloudflare claims containers are necessary for fewer than 10% of agent tasks โ a pointed jab at anyone who built their agent inside a box.
DigitalOcean shipped a public preview of Managed Agents: Firecracker microVMs, per-second CPU billing, 886 milliseconds to start, 305 milliseconds to resume, zero cost while idle. Its Action Gateway hosts MCP endpoints covering 500-plus providers and 16,000-plus tools, with credentials granted at execution time through a proxy and never handed to the model.
And then there is the Blueprint Alliance, announced at Oktane โ not a runtime at all, but a governance layer assembled by twelve vendors and built around MCP, OCSF, SSF and CAEP.
Every one of these is a vendor claim or a forward-looking roadmap, not an independently verified fact. I want to be explicit about that, because the entire infrastructure debate currently rests on numbers supplied by the people selling the infrastructure.
The pricing is the argument
Strip away the marketing and the competitive map becomes almost embarrassingly legible.
Google is selling predictable idle time to individuals. AWS is selling isolation to enterprises running many agents at once. Cloudflare is selling edge density โ three hundred locations beats three regions if your agent lives near the user. DigitalOcean is selling tool plumbing. Okta, through Blueprint, is selling identity.
The billing structures reveal what each vendor secretly believes about agent workloads. Per-second billing with zero idle cost, plus elastic memory reclamation, are two admissions of the same thing: agents spend most of their lives waiting, and peak-provisioned billing was never going to survive contact with them.
Seven days. Fourteen days. Pause-and-resume. These aren't footnotes. They are the SLA semantics of a workload that never had semantics before โ and the fact that all five runtimes disagree on lifecycle length tells you the industry has not settled on what an agent is.
The part the room is missing
Here is where I'll go against the crowd.
Everyone is treating this as a performance race. Latency, cold starts, boot density. But latency was never the real bottleneck in the last cycle either โ I sat through two years of DeFi oracle upgrades that shaved milliseconds off feed times while remaining, structurally, a handful of centralized nodes wearing a decentralization costume. The agent runtime wars are repeating that mistake at a larger scale, and the toll booth is migrating from price feeds to credentials.
Look at the shape of what is being built. DigitalOcean routes 16,000 tools through a single gateway. That is not a feature; it is a chokepoint, and its advertised 99.3% tool-matching accuracy is an internal test figure, which means someone still owns a 0.7% window to call the wrong tool with borrowed credentials. Blueprint bundles identity, security and audit into one reference architecture controlled by twelve members who compete with each other. Cloudflare wants containers demoted to tools so the edge stays the front door.
Nobody is standardizing portability, because portability is the product they are all trying not to sell. Agent state does not migrate between runtimes. Ask what happens to your agent's memory when you move clouds, and you will get a slide about open standards.
What I'm watching into Q4
I am not looking for a winner. I am looking for the seams.
Independent cold-start benchmarks once AWS V2 has been in general availability for a full quarter. Whether Cloudflare publishes container-invocation latency and pricing, or keeps it behind a preview wall. Whether DigitalOcean's tool-matching figure survives third-party red-teaming. Whether Blueprint actually ships open reference code or settles for a logo wall and a press release.
The mixed architecture โ lightweight harness, on-demand containers, externalized state, a tool gateway, and governance bolted across all of it โ is the likeliest end state, and it is already the messiest one. That mess is the opportunity, and it is also the lock-in.
So here is my question for everyone still refreshing charts while these decks drop. If the runtime is just the toll booth, and the identity layer is where the money actually gets collected โ which layer are you buying, and which one are you renting with your entire workload?
We didn't come to BGC to talk about servers. But servers just became the trade.
