The Board Room

Tool servers cost almost nothing to generate; authorising them still costs attention

The most consequential detail in Uber's design is that most of its 800+ tool servers were never hand-written. Each is an MCP server, the emerging standard connector that lets an AI agent call a company system. A workflow called AutoCrawler generates them from Uber's existing API definitions, and an LLM writes the tool descriptions, per Alejandro Saucedo's reporting. Anyone who sat through the last round of internal API sprawl has seen this scene before. A company with a mature API estate can now expose its backend to agents at almost no cost, and the scarce inputs move to governance, identity and context cost.

Uber built three controls into the gateway. Authorization for each tool runs through the company's existing access-control system. Sensitive data is redacted in transit. Every tool is registered disabled until someone explicitly turns it on. That last rule means someone above the team has to own which tools get enabled. A team that auto-generates tools without it has effectively handed its internal API catalogue to callers whose behavior nobody can fully predict. Saucedo also expects agent identity to stay unsolved for years, covering both how a user delegates to an agent and how an autonomous agent stays inside its grant.

The gateway doubles as a hedge against model churn

Uber's gateway translates agent calls into the formats its existing backend already understands, so swapping the model behind an agent never touches backend integration. The design is model-agnostic by construction. A skeptic would call that ordinary abstraction-layer hygiene, and the skeptic is right. It still matters in a week when frontier prices converged and one of them carried an introductory label.

The constraint Uber hit at scale was context bloat: agents hauling too many tool descriptions and oversized responses into every call. The fixes were cost engineering. Agents discover tools incrementally and request only the response fields they need. In a "code mode," coding agents write output to files instead of the model's working memory. The volume of context sent will move agent unit economics more than model selection, and Saucedo supplies the arithmetic: at Argon's post-introductory list price, one maxed-out 1M-token output costs $20.

Where the evidence agrees, and where to discount it

TheSequence's research roundup points the same way. NVIDIA and KAIST's Mid-Harness lifted a small 9B-parameter agent from 50.00% to 68.03% first-try success by having a frontier model check candidate actions before they ran. Microsoft and KAIST's CorpusMap improved enterprise retrieval by 6.45 to 11.74 points while using 43% to 65% of the input tokens. In both cases the gain came from the harness (the checks, routing and data navigation around a model). Architecture Notes reports the same pattern inside Anthropic, where a two-week sprint cut claude.ai's time-to-type at the 75th percentile from 3.1 seconds to 0.55 seconds. The gains came from reliable benchmarks, real-user latency checks and automated regression gates, with humans vetoing complexity that wasn't worth the milliseconds.

Saucedo adds the warning this kind of result always needs. A Google Research and Stanford paper found that harnesses tuned repeatedly against one test set lose much of their gain on tasks they were not tuned for. The paper's regularized method posted the smallest in-house gain and the best score on unseen tasks, 43.6 against 39.7 for the baseline. Any team's harness wins measured on its own evals are best read as partly overfit until held-out data confirms them.

The move

OpenAI's Dots converts this from an architecture question into a deadline, because enterprise access arrives when an admin flips a switch. The sequence that holds up is enablement policy first and gateway second; a company that reverses it will spend the following quarter deciding retroactively which tools should have been on. Uber built in-house because it has Uber's engineering bench. Most companies are better placed to assemble the pattern from open-source and commercial parts: a registry, a gateway, an identity broker (Zalando has open-sourced one) and a disabled-by-default rule.

The durable split is renting the model and owning the gateway, because the gateway is what lets a company switch models without rewiring itself.

Follow the risk, not the capacity

Each financing structure below moves compute risk onto a different party. Read in sequence, they show where that risk ends up.

The last row is the one that reaches the buyer. Microsoft and Anthropic are two of the most sophisticated compute buyers in the market, and both signed with a younger, smaller provider on deals that do not ramp until 2027–2028. A skeptic would say buyers that large can absorb a late ramp. They can. The likelier reading is that financed, available capacity is scarce, and the buyer now carries someone else's build-out risk. If buyers with that much leverage accepted those terms, capacity contracted from a smaller provider on a 2027–2028 ramp probably carries the same construction risk.

Lenders are already marking down borrowed scale

Oracle borrowed heavily to become a major AI data center operator and is down 27% this year, according to The Information. Larry Ellison owns 38% of Oracle and has pledged more than a third of that stake as loan collateral. The same reporting says credit for data center debt is tightening. The Ellison family's media deal shows how fast lenders now react. Skydance's bonds, part of $79B of debt against $12B of EBITDA, fell soon after they began trading and cost investors more than $100M. The Information does not report any delivery shortfall at Oracle. A 2027 plan should still model any debt-funded provider on delivery risk first and price second.

The IPO prices the gap between signed and built

At $35B, Nscale is asking for about 0.34 times its backlog. The Information puts fair value at less than half the target, under 0.17 times. That gap is the market's discount for delivery and financing risk. Spread across 12 unbuilt sites, the backlog works out to a crude $8.6B of contracted revenue per site. Each site has to clear financing, power, construction and chip supply before it earns anything. Rising debt costs have to land somewhere. The provider can pass them through as higher compute prices or absorb them as thinner margins. When neither works, projects stall.

A second constraint may be forming. SRE Weekly flags Gergely Orosz describing an emerging CPU shortage, with no detail on cause, size or region. The report is thin enough that it is worth checking with cloud account teams before pricing it in. If it holds, efficiency becomes growth capacity. Cloudflare has saved another ~100TB of RAM through engineering, and Netflix automated container right-sizing with built-in fallback. Both companies can grow on hardware they already own.

The move

The practical response is to sort every 2027–2028 commitment into three tiers: capacity that is operating, capacity that is financed but not yet running, and capacity that is contracted but not fully financed. The first counts as supply and the second counts with a schedule buffer. The third is an option, and it gets hedged like one. The terms then need rewriting so payment moves only when the provider's financing does. The tradeoff is renegotiation friction this quarter against stranded exposure in 2027. If debt keeps tightening, cash held in reserve becomes useful, because unfinanced projects and stranded contracts can reach the market on better terms than today's.

Contracted capacity that is not fully financed belongs in the 2027–2028 plan as a hedged option, with payment tied to the provider's financing.

How a quiet exit becomes a short thesis

A firm called Canary Data tracks company leadership pages and LinkedIn profiles to spot departures that were never announced. Short sellers then pair those exits with operational bad news. Tesla shows the sequence. Four senior leaders left without disclosure: the 10-year head of Optimus robot sensors, the Dojo/Autopilot AI hardware director, the Texas Gigafactory quality head and the Cybercab new-product-introduction head. Within days, The Information reported that Tesla is struggling to scale Optimus manufacturing. Each fact alone is routine. Grouped by someone with a position, they become a thesis about execution.

The lesson is not that executives leave. It is that silence no longer buys time. Staying quiet simply hands the timing of your story to someone else.

Who is setting the price for your leaders

TheSequence reports that Meta hired MongoDB's sitting CEO, CJ Desai, to run its new enterprise platform. A big-tech company recruiting a public-company CEO ten months into the job shows the AI talent war has reached the C-suite and its direct reports. MongoDB paid the cost. It had no ready successor and brought back the predecessor who ran it for 11 years as interim CEO. For MongoDB's customers, that is a vendor-continuity question. For you, it means companies with deeper pockets now set the price for your AI, data and infrastructure leaders.

The same scraping reaches your claims

The method goes beyond org charts. Hunterbrook counted agent microsites on Ethos's own web domain. It found about 70% of roughly 178,000 agents sit inside three multilevel-marketing agencies, though Ethos markets "algorithmic" life insurance. J Capital found that Insperity's projected profit recovery depends partly on capitalizing costs it used to expense, while its new HRScale platform has only a few beta clients. If you sell "AI-native," your actual go-to-market has to survive someone scraping your website.

Weigh the source. The Bear Cave is owned by Hunterbrook Media, and its affiliate Hunterbrook Capital was short Ethos when this was published. Treat that thesis as informed advocacy until it is independently confirmed. The scraping method works no matter who uses it.

There is an opening too. Tesla's quiet losses in robot sensing, AI hardware and new-product introduction create an unusual recruiting window for anyone building in robotics, autonomy or custom silicon.

Assume everything public about your company will be scraped, grouped and narrated by someone with a position, so announce your own changes first.