The week AI started eating your software budget

We have left the phase where AI capability just shows up for free and impresses everyone in the demo. Capability is still spreading, and in some places it is getting genuinely cheaper. But this was the week the bills arrived, all at once, in budgets, in authority, in infrastructure, and in trust. Not a doom week (even if headline algorithms push that to the top). A growing-up week. Below is what actually moved, and how I would read it if you have to make decisions with real money and real people behind them.

Models, Labs & the Capability Frontier — advanced. Two open-weight releases and one delay tell the story. Moonshot shipped Kimi K3, a very large open-weight model that lands right near the frontier. K2 already had a strong reputation, and K3 pushes further. It is lumpy, meaning it is top-tier at some tasks and noticeably weaker at others, and it is genuinely large, so you are not running this in a home lab any time soon. (the back of the envelope says you would need 16 nodes of DGX Spark w 128GB RAM each… so yeah, not a home use model) But it is open weights, which means people can train off it, distill it, and adapt it. It’s also currently available at 1/10th the cost of the less lumpy frontier models which is the real headline mover. Thinking Machines released Inkling, and if you are looking for US open-weights models it may be the more interesting of the two, because it is not chasing benchmark supremacy at all. It is built for fine-tuning on your own company data, so the buzz crowd who ranked it and shrugged mostly missed the point in my opinion. I think it’s the quiet headline that more business people should be paying attention to. Meanwhile Google's Gemini 3.5 Pro is reportedly still held back for falling short of internal targets, especially in coding. (ouch)

My read, through Capability-to-Consequence: the model is becoming raw material, and the frontier is not a straight line. When advanced capability comes in an open-weight format you can adapt, the scarce, valuable work moves downstream into adaptation, evaluation, trust, and cost per completed task. And the Gemini delay is the humbling companion note. This is not "Google is doomed." Google has the hardware, the software, the research depth, and a real multimodal edge, and will be in this race for a very long time. The lesson is that even the best-resourced lab on earth cannot reliably convert spend into the next step-change on a schedule. Coding matters here specifically because it underpins agents and enterprise automation, so a coding shortfall is not cosmetic. What to do: if your AI strategy assumes the next model release arrives on schedule and solves your problem, you do not have a strategy, you have a hope. Separate ecosystem strength from release certainty, and build the adaptation and evaluation muscle now, because that is the part you actually control.

Workflows, Not Models — advanced. OpenAI proposed a scorecard that measures successful completed work rather than seats, logins, or adoption. I pause a little, because it is a touch self-serving. Their framing nudges you toward "just use the best model, it one-shots everything." Sometimes true, often not, and the cost-performance math for your specific task is exactly the thing they gloss over. But the core idea is right, and it is the drum I have been beating for a while. Through the Systems-Over-Models lens: value lives in the outcome, not the logo on the model. What to do: every company should run its own evaluation ecosystem where you test models against your real tasks, measure the return, and stay comfortable that you are getting value you can see. A cheaper model that clears the bar is a better business decision than a prestigious one that does it slightly faster. And do not outsource the scorecard to a vendor whose revenue depends on the answer.

Compute, Capital & Power — accelerated. This is the most operationally practical cluster of the week, and it moved hard. IBM warned that customers are shifting spend toward servers, storage, and memory to run their own AI, and that software budgets are shrinking as a result. Several large software deals reportedly failed to close, and the warning set off a broader software-sector selloff. On the supply side, Samsung is consolidating around Texas: moving toward next-generation Tesla AI5 chip production in Taylor, relocating US headquarters functions out of New Jersey with 739 roles affected, and trimming around a hundred more in Plano as consumer electronics get squeezed by the same rising chip costs that are lifting the chip business. And on the permission side, New York imposed a one-year statewide moratorium on new hyperscale data centers to sort out ratepayer, community, and environmental standards, while Australia stood up a central AI office with planned rules pushing data centers to limit water use and even become net energy producers.

Through Primary vs Secondary Effects: the primary effect everyone sees is AI demand lifting chips and infrastructure. The secondary effects are the real story. AI investment can create a boom in one division while forcing restructuring in another inside the same company, as Samsung shows. And "we want domestic AI capacity" in the abstract collides with "who pays for the power, water, and grid" the moment a data center shows up in an actual town. The cloud has a zip code now, and it needs local permission to build. What to do: if you sell software, expect harder renewals and bring outcome data to every one of them. If you buy it, this is a strong negotiating season, though watch out for everyone bolting "AI" onto the same product at a higher price. And if your roadmap depends on compute in a specific region, start treating local politics as a supply-chain risk.

Safety, Trust & Regulation — accelerated. Developers reported GPT-5.6 Sol taking destructive actions, including deleting files and production data, and doing work beyond what it was scoped to do. I have felt the milder version myself: models, Fable included, cheerfully adding sections to my own systems that I never asked for because they now overthink things, which I then have to go back and rip out. The flip side is Anthropic's permission-hungry (starved?) style, where I wake up to a job that sat patiently all night waiting for approval instead of running in the tools and box that I gave it because it thought of a better way. Annoying in the moment, but harder to say it’s not the right instinct. (Getting it to respect the lanes given and permitted vs. asking before going out would be even better.) Separately, leaders at Google DeepMind, OpenAI, and Anthropic have signaled public support for stronger third-party frontier testing, even as employees at those same companies fund competing governance approaches via political contributions, which tells you this is not a settled, one-size-fits-all consensus.

Through the Translation Layer lens: intelligence without authority boundaries is a liability. A chatbot can be wrong. An agent can be wrong with permissions… in production… at speed. That is a categorically bigger problem, and it means permission scope, staged execution, reversibility, and audit trails are now core agent features, not nice-to-haves. We solved this with people using management structure, not smarter employees (there is probably a study somewhere about how smarter people chafe at bureaucratic control, no wonder smarter AI does too). Same answer here. What to do: before you widen any agent's production access, write down what it may touch, what happens when it exceeds that, and how you roll it back. The most expensive AI error you will ever have is a correct-looking action taken with the wrong authority. Third-party oversight also gets more credible precisely because the labs themselves are acknowledging shared release risk.

The Human Layer — advanced. Former Meta employees filed suit alleging the company used AI-assisted criteria in its layoffs, and that workers with medical conditions, disabilities, or pregnancy-related leave were disproportionately affected because they showed up as “less productive” in the data, including token-usage boards. Meta's line is that humans made every decision. These are allegations, not proven facts, and I want to be careful with the wording. But watch the mechanism, because that is what matters even before any verdict. A human can sign off on every decision while an opaque system quietly structures the ranking that human is approving. Medical leave and disability status get indirectly encoded through productivity metrics without anyone typing them in. If this case proves out, it becomes a defining example of algorithmic management, productivity measurement, and employment law colliding, and it is exactly the kind of story that lights up every anti-AI concern people already carry. If you run productivity analytics anywhere near workforce decisions, especially in a health system with real medical-leave data, this is your warning shot: build a way for people to challenge algorithmically influenced calls.

Innovation & Health/Science — advanced. For a health system, this is the section that should make you lean forward. Researchers tested a hybrid neural bypass in one participant with chronic complete tetraplegia, linking a brain-computer interface with muscle stimulation plus spinal and cortical stimulation. The participant was able to self-feed and manipulate delicate objects, and here is the part that genuinely excited me: some motor and tactile improvement persisted even when the system was switched off. It is peer-reviewed and real, and it is still a single participant, so temper the excitement with the sample size. But the important thing is that it’s not sci-fi or magic. It is a system that helps the person act now while helping the body relearn over time, combining decoding, stimulation, reinforcement learning, and neuroplasticity into one integrated whole. Separately, researchers used AI protein design to create synthetic TnpB genome-editing enzymes (CRISPR) with sequences found nowhere in nature, reportedly editing more efficiently than natural counterparts and cutting unintended edits sharply. AI is moving from analyzing the proteins we found to inventing tools nature never made. The future of gene editing with precision and no side effects is moving out of science fiction territory. (Standard caveat of this is still in research and not yet here of course)

Through Capability-to-Consequence: the strongest AI-enabled health advances are not single models, they are integrated systems that validate, control, and translate capability into a human outcome. That is the whole thesis of this show in one hopeful example. What to watch: both results still need larger studies and extensive safety validation before they touch a patient, so treat them as direction, not delivery.

Embodiment & the Physical World — held. The neural bypass touches physical restoration, but there was not enough independent robotics movement this week to make embodiment a lead theme.

The connective read: capability is spreading, but advantage now depends on budget discipline, authority boundaries, infrastructure strategy, and trust. The next phase of AI will not be won by capability alone. It will be won by the systems that turn capability into value without losing the trust required to keep using it.

One question: when your AI agent takes an action that is correct but that you never authorized, do you count it as a win or an incident? Where you draw that line says a lot about how ready your organization actually is. Hit reply, I read them.

If you want the live version, with the visual board and the takes as they land, the episode is the companion to this issue on YouTube

Chuck
Chief Technology Troublemaker chuckdevries.ai

Someone forwarded you trouble.
They clearly think you can handle it. Trouble Worth Making is where I work out loud on how to stop bolting AI onto a pre-AI company and actually rebuild the operating model underneath it. One issue every week or two, no filler, the occasional uncomfortable question.

If the person who sent this has good taste (obviously, they do), subscribe here and get the next one straight from the source.