Token Budget Wars, Software After AI, Defensibility Fallacy

This week's TLDR Founders digest is a coordinated shot across the bow of three pieces of conventional wisdom that the AI era is breaking. Token budgets are hitting corporate walls faster than anyone planned. The "software era" is ending and the "harness era" is starting. And the obsession with pre-launch defensibility is, for most startups, a trap. Read together, these three stories are a coherent picture of where value is moving in the AI value chain — and it's not where most pitch decks say it is.

What You Need to Know: Uber capped engineer AI spending at $1,500/month after burning through its entire 2026 AI budget in four months; Tomasz Tunguz (Theory Ventures) published "Software After AI" arguing the next era of software is the "harness era" built on seven disciplines (context, tools, orchestration, state, sandbox, observability, cost optimization); and Hawkhill Ventures published "The Fallacy of Defensibility" arguing that most great startups don't start with moats — they build them through years of execution.

Why It Matters

• AI spend is being rationed at the company level, not the user level. Uber's $1,500/month cap and Microsoft's cancellation of Claude Code licenses are early signals. If your revenue model assumes unlimited per-user inference, the procurement model is changing under you.
• The "harness" is the new SaaS. Tunguz's framing: AI outmoded fixed-workflow SaaS the same way the internet outmoded print. The companies that build the seven layers of the harness — context, tools, orchestration, state, sandbox, observability, cost — are the ones that will dominate the next decade.
• Model access is not a moat, and never was. Hawkhill's argument: defensibility is a long-tail result of execution, not a starting condition. "What stops OpenAI from building this?" is a useful question, but it's a 5-year question, not a seed-stage question.
• Token-to-outcome tracking is the new metric. The shift is from "are people using it?" to "what business outcome did those tokens buy?" If you sell AI to enterprises, your next renewal conversation will be about cost-per-resolved-ticket, not active users.
• The harness layer is the defensibility layer. Model gets commoditized; the harness doesn't. The companies that own context, tool registries, and orchestration in a specific vertical win because the model is interchangeable and the harness is not.

What Actually Happened

The Token Budget Wall

Uber's Chief Technology Officer revealed in April that the company had burned through its entire 2026 AI budget in four months, after leadership encouraged staff to "use AI as much as possible" and ranked internal usage competitively on leaderboards. By June 2, Uber had instituted a $1,500 monthly cap per employee per agentic coding tool, applicable to Claude Code, Cursor, and similar products. The cap is trackable via an internal dashboard and can be exceeded with manager permission in certain cases.

Uber COO Andrew Macdonald was more pointed in a May podcast appearance: "It's very hard to draw a line" between AI usage and new consumer features. The blunt version: Uber isn't sure its AI spend is producing measurable value.

The pattern is industry-wide. Microsoft cancelled Claude Code licenses after seeing the price tag. GitHub Copilot's switch to token-based billing in late May 2026 sparked user revolt. Bain's 2026 survey found that AI "delivers less cost reduction than many firms predicted." The WSJ ran "Corporate America Is Starting to Ration AI as Costs Skyrocket" in late May. The TLDR framing — token-to-outcome tracking: cost per resolved ticket, processed claim, reviewed contract, avoided hire, or dollar of revenue moved — is the new procurement language. (Source)

Software After AI: The Harness Era

Tomasz Tunguz (GP at Theory Ventures) published "Software After AI" on May 27. The thesis: the end of the software era is the beginning of the harness era. "AI outmoded SaaS managed databases with fixed workflows with intelligence. Like a mustang, AI is powerful but wild. Harnessing the power means domestication."

The harness is built on seven disciplines, each a startup category in its own right:

1. Context & memory — bespoke retrieval for each use case. A radiologist's context system is not a paralegal's. The "context database" is the recipe book of how each business actually runs. 2. Tools & action — a tool registry with argument validation, gated sensitive actions, and MCP as connective tissue. The quality of a harness depends on how many tools it can safely expose and how cleanly it handles failures. 3. Orchestration & loop — the think-act-observe-repeat agentic loop with planning, decomposition, sub-agents, retries, and stop conditions. Closed-loop patterns that learn from each run will differentiate vendors. 4. State & persistence — when a harness crashes at step 7 of a 10-step task, it should resume at step 8, not restart. File systems, checkpoints, session threads, artifact storage. 5. Sandbox & compute — isolated Unix workspaces, controlled network egress, credentials living outside the model. 6. Observability & governance — tracing, logging, evals as regression tests, humans in the loop for high-stakes decisions. 7. Cost & workflow optimization — architectural judgment on what should be deterministic vs. non-deterministic, which model tier to use per step.

Tunguz's punchline: "What happens when every company has access to the same model? The best riders win." (Source)

The Fallacy of Defensibility

Hawkhill Ventures (a seed-stage firm) published a short essay arguing that the "what stops OpenAI from building this?" question, when asked at seed, is mostly a trap. The argument: "The argument about moat and defensibility for 9.9 out of 10 early-stage startups is ridiculous because almost no company starts with a moat. Most great businesses grow into defensible positions."

The case studies are pointed. Ramp did not invent corporate cards. Stripe did not invent payments. Toast did not invent restaurant software. Datadog did not invent infrastructure monitoring. Each entered a crowded arena and won through years of execution — better software, deeper customer trust, better distribution, better hiring, better culture. The moat was the result, not the prerequisite.

The punchline: "Strategy eventually gets copied. Features get copied even faster. Models and infrastructure are becoming commoditized at a pace we have never seen before. What remains difficult to replicate is a team that consistently executes at a high level over many years."

This isn't an anti-moat argument. It's an anti-asking-the-wrong-question argument. Moats matter; they just don't show up at seed. (Source)

The Take

Read these three together and a single picture emerges: the value is moving down from the model and up from the workflow.

The token budget war tells you the model is becoming a commodity line item, rationed like cloud egress. The harness essay tells you the seven layers above the model are where the durable value lives. The defensibility essay tells you the moat in those seven layers is built by years of execution, not by a clever founding insight.

For builders, the immediate consequences are concrete.

If you're selling AI to enterprises, change your pricing and packaging immediately. Per-seat subscriptions that assume unlimited token usage are about to be renegotiated. The procurement question is going to be: "What did your tokens buy?" If you don't have a clean answer, you're going to lose the renewal. Build the token-to-outcome tracking now, even if your customer doesn't ask for it. They'll ask in Q4.

If you're building an AI startup, stop pitching the model. Pitch the harness. The model is the same model everyone else has. The context, the tool registry, the orchestration, the state management, the sandbox, the observability, the cost optimization — that's the company. Tunguz's seven categories map to seven startup sub-categories, and the companies that win in each are the ones that own the vertical-specific context database. "Context is the new code" is the aphorism to internalize.

If you're a seed-stage founder in a "commoditized" category, don't apologize for not having a moat. The moat comes from the next five years of execution. Spend the seed round hiring the team that can execute for five years, not the team that can build a moat in 18 months. Ramp didn't win corporate cards by having a moat. They won by hiring people who gave a shit about finance teams for a decade.

The piece I'd push back on is Hawkhill's implicit timeline. Five years of execution is what it took Ramp. Most seed investors want to see a moat in 24 months. The gap between what the essay says and what the term sheet requires is the actual tension in the market. Founders who can sell the "we will execute for five years" story credibly are the ones who will get funded at premium terms in 2026. The rest will be told to come back when they have a moat.

Quick Summary

Uber's $1,500/month per-engineer AI cap is the leading edge of enterprise token rationing; Tomasz Tunguz's "Software After AI" argues the next era is the harness era (seven disciplines: context, tools, orchestration, state, sandbox, observability, cost); Hawkhill's "Fallacy of Defensibility" argues most great companies don't start with moats — they build them through years of execution, citing Ramp, Stripe, Toast, and Datadog.

Sources

• TechCrunch — Uber caps employee AI spending after blowing through budget in 4 months
• Tomasz Tunguz — Software After AI
• Hawkhill Ventures — The Fallacy of "Defensibility"
• TLDR Founders — Token Budget Wars, Software After AI, Defensibility Fallacy (2026-05-29)
• Bloomberg via TechCrunch — Bain survey: AI delivers less cost reduction than many firms predicted