Today’s infrastructure news keeps running into the same question: who should own the agent stack, and what do you give up when you let someone else? Cloudflare has the most numbers, and they cut both ways.
1. A forecast says 7 in 10 enterprises will drop vendor agents
The Register’s headline says 7 in 10 enterprises are expected to abandon vendor-built agentic AI by 2028. It’s a striking number, but the page text supplied for this briefing is only navigation links. We can’t see who made the forecast, how it defines “abandon,” or what replaces the vendor tools.
That gap matters. “Abandon” could mean a full rip-and-replace, a switch to a different vendor, or pulling a few workloads in-house. A forecast is also not measured behavior, and predictions three years out are cheap to make.
Founders selling agents should still take it seriously, because it attacks the pitch of buying ours instead of building your own. Read the underlying report before repeating the statistic in a deck.
2. Cloudflare’s cheaper model router trades accuracy for savings
Cloudflare’s Auto Router picks a model per request, and its own benchmark shows the catch. On an internal knowledge-work test of 97 tasks, cloudflare/auto succeeded 86.6% of the time at $0.0084 per success. Claude Opus 5.5 hit 96.6% at $0.0210. GPT-6 Sol landed at 84.2% and $0.0108. So the router is clearly cheaper than Opus and modestly better than Sol, but it fails more often than Opus, 39 of 291 trials against 10. Cloudflare ran the test itself and calls the router a public beta.
The other launches say where Cloudflare is heading. Containers now let code choose each sandbox’s image at runtime, and ComputeSDK’s benchmark put median startup at 648 milliseconds, down from 4.049 seconds. Cloudflare’s registrar can be driven by agents through an API, MCP, and a CLI, with a sandbox for testing.
The pattern is that Cloudflare wants to be the control plane for agents. That is convenient, but every primitive here, from Durable Objects to AI Gateway, is also a dependency. Building your own stack on Cloudflare is still building on someone else’s.
Cloudflare Blog · Cloudflare Blog · Cloudflare Blog
3. OpenAI built the app store storefront but skipped the cash register
OpenAI’s Dev Day plan turns ChatGPT into a place where software gets discovered, launched, and used. According to TechCrunch, it did not announce a billing or revenue-sharing system like the economic layer of Apple’s and Google’s stores. That is an odd omission for something pitched as an app store rival.
The pieces so far are in-conversation app suggestions, interactive panels, and “Sign in with ChatGPT,” which lets users bring their AI allowance to 16 launch partners. There is also an enterprise marketplace with 30-plus partners, including Adobe, Figma, Salesforce, and Baseten. Eligible customers can put part of their OpenAI commitment toward approved partner software.
That last feature is the AWS Marketplace move The Register’s headline points to. My read is that it gives partners a procurement path and keeps enterprise spend inside OpenAI’s contracts.
The Register also calls Anthropic “stubbornly Microsoftian,” but the supplied text doesn’t include that argument. The open question is whether OpenAI eventually takes a cut.
4. FedEx is both investor and customer in Harbinger’s $300M order
FedEx’s $300 million order for 2,000 Harbinger electric trucks works out to roughly $150,000 per truck. It also comes from a company that led Harbinger’s $160 million Series C late last year. TechCrunch reports that Harbinger has delivered 53 trucks so far and promises all 2,000 by the end of next year.
That is a steep ramp for a startup founded in 2022. It is also a neat piece of validation that comes partly from an investor. The order shows demand from one large customer, not proof of unit economics, and the source doesn’t discuss margins.
Harbinger’s single-chassis pitch is also widening. It is moving into hybrid emergency vehicles, battery packs sold for energy storage, and defense, and it acquired an autonomous driving company in February. Axios reported in May that it is considering an IPO. The test for an IPO story is whether the diversification reflects demand or a search for more revenue.
5. Restate’s $20M bet: agents need durable execution, and Temporal is the target
Restate raised a $20 million Series A, not a seed round, and it arrives in the shadow of Temporal’s $550 million Series E at a $12.55 billion valuation earlier this month. The Berlin startup, led by Apache Flink co-creator Stephen Ewen, says it landed multiple six- and seven-figure contracts in recent months. Singular led, with Redpoint and Capital One Ventures participating.
Ewen’s own account is candid: Restate “never was built for agents in the beginning” but matched the problems they raise. Agent workflows run longer and take less predictable paths, so recovering mid-process matters. Restate’s differentiator is its own storage, replication, and redundancy layers instead of an external database, which Ewen says keeps it fast and lightweight. Replit is a named customer.
Temporal’s raise shows investors already see the category as large. Restate is betting that a leaner engine can win workloads Temporal treats as heavyweight.
