The week the AI budget came due
Uber torched a year of AI budget in four months, MIT still says 95% of pilots show no return, OpenAI built its own chip, and Washington began handing out frontier models by name.
Week of June 22-28, 2026 · by the Hotovo AI team
TL,DR
- The bill came due. AI capex is enormous while measurable returns stay rare. Uber burned its entire 2026 AI budget in four months, and its COO still cannot tie the spend to a single better product.
- MIT's number hasn't moved. Roughly 95% of enterprise generative-AI pilots show no measurable business impact, and the cause is workflow and integration, not model quality.
- OpenAI built its own chip. Jalapeno, its first custom inference chip with Broadcom, was taped out in nine months and is aimed at cheaper inference and less reliance on Nvidia.
- Washington started handing out models by name. GPT-5.6 Sol shipped to about 20 government-approved partners, the first US frontier model released behind a government-managed access list.
- Intelligence keeps getting cheaper. Open-weights GLM-5.2 lands near frontier quality at a fraction of the cost, and Apple is putting capable models on-device this fall.
- Spicy pick: Nvidia, Google and SpaceX are all racing to put AI data centers in orbit to escape land, power and cooling limits on the ground.
Inside Uber this spring, engineering teams were climbing a leaderboard. The metric was not features shipped or bugs fixed. It was tokens consumed. The more AI a team used, the higher it ranked, and within four months the company had spent its entire 2026 AI budget. By spring, about 95% of Uber engineers were using AI tools and roughly 70% of committed code was machine-generated. Then the COO, Andrew Macdonald, said the quiet part out loud: he could not draw a clear line between all that spending and one better product for riders. That admission, reported by Fortune, is the AI economy in miniature, and this week it stopped being an anecdote and became the theme.
The main story: the AI budget came due
For two years the AI conversation ran on capability, with bigger models and longer context windows. This week it turned into an accounting question. MIT's NANDA study still finds that around 95% of enterprise generative-AI pilots deliver no measurable impact on the bottom line, and its researchers are blunt about why: the problem is the learning gap inside organisations and weak integration, not the quality of the models. More than half of generative-AI budgets went to sales and marketing, while the clearest returns sat in unglamorous back-office automation.
Set that against the spending. The four largest US technology companies have hundreds of billions of dollars in AI capex planned for the year, and KPMG put the average enterprise AI investment near 269 million dollars over the next twelve months. The distance between what companies spend and what they can prove is now the defining enterprise-AI story of 2026.
Why it matters - the Hotovo read
AI still pays, but the payoff lives in the engineering around the model rather than the model itself. The teams seeing a return designed the workflow first, measured it against real business criteria, and put AI where it actually removes work. That is how we approach it at Hotovo: AI integrations are not magic, they are engineering. When we built g-Xperts' AI copilot to automate heavy business reporting, the value came from the plumbing and the evaluation, not a demo. We check generated output at runtime against business criteria with tools like Promptfoo and Echo before anyone relies on it, and we run private code review inside the customer's own Azure environment so proprietary code never leaves it, all governed under ISO/IEC 42001. A token leaderboard measures activity. A workflow that removes an hour of reporting measures value.
Also worth your attention
OpenAI built its own chip
OpenAI and Broadcom unveiled Jalapeno, the company's first custom inference chip, taken from design to fabrication in about nine months with OpenAI's own models helping build it. Early testing shows performance per watt the company calls substantially better than the current state of the art, with deployment at scale due late 2026 and reports of roughly half the inference cost. Google already has its TPUs and Amazon its Trainium; now the model maker makes the silicon too. The signal underneath the news is simple: whoever controls the cost of inference controls the economics of everything built on top, which loops straight back to the return question above.
Washington becomes AI's gatekeeper (a follow-up)
Two weeks ago we covered the US shutoff of Anthropic's Fable 5 and Mythos 5. This week the same dynamic widened. OpenAI released GPT-5.6 Sol to roughly 20 partners whose names were individually approved by the government, the first time a US frontier model has shipped behind a government-managed access list, pending a formal review process targeting August. OpenAI said plainly that it does not believe this should become the default. What changed since the ban is the direction of travel: from blocking a model to allocating one.
The Hotovo read on the drama
When access to a model can change by memo, resilience has to be designed in from the start. We build AI apps, agents and automations to be model-portable, with a primary model, a secondary and an open-weights fallback, plus abstraction layers so components swap without a rewrite, and governance under ISO/IEC 42001 and ISO/IEC 27001. A provider being blocked, rate-limited or recalled should never take a customer's product down. That is the whole point: we protect our customers' uptime first.
The price of intelligence keeps falling
While frontier access turns political, capable AI keeps getting cheaper. Z.ai's open-weights GLM-5.2 landed near the leading proprietary models on agentic and coding benchmarks at roughly a quarter of the cost, and Apple confirmed on-device foundation models, built with Google, shipping this fall. For most business work, good-enough intelligence is now nearly free, which makes the return question sharper still. If the model is cheap, the failures are about how you use it.
Spicy pick: data centers are heading to orbit
The least mainstream story of the week is also the most audacious. Nvidia, Google and SpaceX have all moved toward putting AI compute in space. Nvidia launched space-ready platforms for orbital data centers, and Google is paying SpaceX for bridge GPU capacity. The logic is physical: on the ground, land, power and water cooling are the real bottlenecks, whereas in orbit the sun never sets and the vacuum is free cooling. It is early and unproven, but it tells you how hard the compute crunch has become.
AI tip of the week: route your models
Stop sending every request to the most expensive model. Set up simple model routing: let a cheap or open-weights model handle drafts, classification and bulk work, and escalate to a frontier model only for the final or highest-stakes step. Keep one open-weights model you can self-host as a fallback. You will usually cut cost sharply, and, as this week showed, you will not go dark if a provider gets blocked or throttled. It is one of the quickest ways to make your AI budget go further.
The bottom line
The hype cycle is quietly turning into an accounting cycle. The winners in the second half of 2026 will not be the companies that bought the most tokens. They will be the ones that designed the workflow, measured the outcome, and built so that no single chip, model or memo can knock them offline. Spend less on demos, and engineer more for value.
Sources
Newsletters (week of June 22-28, 2026): The Deep View, LaunchpadFast, The Batch (DeepLearning.AI), This Week in AI Club, AI Valley, The Neuron, Exponential View.
Verification and context:
Fortune - Uber's 2026 AI budget
Fortune - MIT report: 95% of GenAI pilots failing
TechCrunch - OpenAI unveils its first custom chip, built by Broadcom
TechCrunch - OpenAI limits GPT-5.6 rollout after government request