The story is not that production got cheap. It is that the scarce thing moved, and nobody agrees yet on where it moved to.

💡 Top tip for this week's roundup - read the TL;DR, then jump to the section closest to your current problem. Turner and SaaStr are the two to read together: one is the company doing the cutting, the other is the vendor being cut.

Remember, we've built an ever-growing library of our top podcast summaries. Check it out here.

Featured in this round-up:

🕒 Estimated reading time: 10 mins. Time saved: 14+ hours! 🔥

Not your topic this week? Try these five instead:

———————————————————————————

Start with the cleanest demonstration of the week, because it comes with an invoice.

Fred Turner is co-founder and CEO of Curative. The headline story is the one 20VC put in the title: a COVID testing operation that went from $0 to $5BN in revenue in three years, grew from 7 to 7,000 employees in nine months, and ran 206,000 tests in a single day at its December 2020 peak. Turner then pushed roughly $500m of those profits into building a health insurance business, now valued at around $1.3bn.

Good story. The back half is the one that matters. Asked whether he buys the "SaaS is dead" theory, Turner answers with his cancellation list: an internal CRM, vibe coded in two months, that "is working better, that is managing our process better... And no one was using Salesforce anymore." $600,000 a year. "Gone to zero." He is straight about the catch - "the maintenance is definitely one of the most challenging pieces" - but reckons that for any business of reasonable scale, several FTEs babysitting "this archaic software platform" becomes one good engineer building features on request. Around 80% of Curative's SaaS spend goes this year, tracked in a meeting with a slide listing which contracts are due and whose job it is to tell them. What survives is infrastructure: he names Sentry, and admits Slack has been "notoriously hard" to leave.

Then the sharp bit. About 45 people negotiate Curative's provider network contracts. Earlier this year they launched an agent called Gwen that runs the whole workflow - researches the practice, checks transparency files for what other payers pay them, finds the email, chases until someone replies.

Stebbings asks what happens if Anthropic doubles its prices. Fine, says Turner. So would 5x. Which is exactly why their Anthropic bill has "six X'd every month" for six or seven months, from tens of thousands to millions: "we just keep finding new things to do with it." Meanwhile credentialing - checking every doctor's licence and malpractice history, two to three months at about $50 a time - was "the first one that went to zero people". Curative did not buy a cheaper CRM. It stopped being a customer for that category, and the saving showed up as a vendor losing a logo and never being told why.

Key takeaways:

Now the same event from the other side of the table.

SaaStr 869 makes the case that internal AI agents are quietly becoming your most dangerous competitor - not by outselling you, but by inspecting you. An agent with access to your APIs sees how thin the surface really is, works out it could rebuild the useful 20%, and does it over a weekend. Your first signal is a non-renewal with no explanation attached. They name names on the receiving end - HeySummit and Squarespace have both lost seats this way - and are refreshingly candid about being perpetrators too: their own AI VP of Marketing, built in Replit on their existing stack's APIs, rebuilt a $4,000 tool in about an hour. They moved ten years of history off Marketo for $14.

Put that next to Turner and the loop closes in a single week - SaaStr naming stealth churn as a phenomenon, Curative doing it with the invoice to prove it. If you sell software, your renewal conversation may already have been lost inside a customer's coding environment months before anyone told your CRM. Their prescription is blunt: your agent has to proactively teach customers everything it can absorb for them, because if it does not, a competitor's agent will happily absorb your function instead.

Key takeaways:

The floor fell at the infrastructure layer too, and this one happened in public.

On 16-07, Moonshot released Kimi K3: a sparse mixture-of-experts model at roughly 2.8 trillion parameters with a 1M-token context window, the largest open-track model so far. Pricing landed at $3 per million input and $15 output, open weights promised for 27-07. It beat Fable 5 and GPT-5.6 Sol on front-end coding in Arena evals, triggered a sell-off across AI names, and had everyone reaching for "DeepSeek moment" again. The pricing matters more than the benchmark - Grok 4.5 sits at $6 output, Gemini has been reported as low as $1.25 input. Frontier-adjacent capability is drifting toward commodity pricing faster than most 2026 plans assumed.

One contested claim underneath it all, which should stay labelled contested: Anthropic has accused Moonshot and other Chinese labs of industrial-scale distillation, training on millions of exchanges with US models. Unresolved - and how you weigh it decides whether Kimi K3 reads as real competition or as a derivative.

Worth carrying one more angle into that argument. The sharpest counter-take doing the rounds is that "American labs versus Chinese labs" is itself a commercial position - the real threat to closed frontier labs being open models generally, of which the best currently happen to be Chinese. "Do you really want to use their model" sells to a CIO far better than "do you really want to use a free one". Both can be true at once.

Key takeaways:

If you read one section this week, make it this one.

Dex Horthy is CEO of HumanLayer and the person who coined "context engineering". He got there the expensive way. In July 2025 his team ran what he calls a dark factory: the model writes the code, humans review nothing. Four months later they threw the entire system out. Production broke, no amount of prompting would get Opus 4.1 to find the root cause, and days of wading through spaghetti code turned up a primary key wrongly routed through the whole codebase. The fix was the cheap part. The expensive part was three weeks re-onboarding humans to a codebase no human had ever read.

His diagnosis of why is the part to steal. Coding models are optimised for SWE-bench-style benchmarks that reward reproducing a known fix in a known repository - but you cannot evaluate bad architecture with a unit test. The thing that quietly rots is the exact thing the benchmark is blind to. He also thinks the window is shrinking: four months to wreck a codebase in 2025, less now that generation is faster.

The technique is worth stealing wholesale:

Key takeaways:

Rory O'Driscoll of Scale Venture Partners has been investing in software for over thirty years, which makes him a useful counterweight to a week of collapse stories. His position: software is not dead, it just got a lot harder to win.

The number framing everything else is roughly $688BN of hyperscaler AI capex in 2026. His analogy does the work. Microsoft owned client-server. AWS owned cloud. Foundation model providers will own their layer - and in every previous cycle, plenty of software companies still won by building on top of a layer somebody else owned.

That turns a vague worry into four sharper questions. Which layer are you actually on? What moat are you building there? Is your growth real or borrowed from the current wave? And does your product sell itself?

Read alongside Turner and SaaStr, this is the corrective. Curative did not stop buying software. It stopped buying that software and moved the budget one layer down to a model provider. The spend did not vanish, it relocated.

Key takeaways:

The most striking number of the week comes from outside software entirely.

Andy Beam and Rafa Gómez-Bombarelli of Lila Sciences argue that science - not the internet - is the last genuinely untapped source of training data, and that laboratories should be built like data centres. Instruments become nodes on a graph, connected by a magnetically levitating transport layer they describe as a physical PCI bus, with experiments orchestrated the way a scheduler works a compute queue.

Then the number: two to three people built a complete in vivo CAR-T therapy in about six months. Capstan Therapeutics took roughly five years and over $100m to reach comparable ground, and was acquired by AbbVie for $2.1bn in early 2026.

Gómez-Bombarelli stops this being a straight acceleration story. The rigour you apply to an AI-generated hypothesis cannot be lower than the rigour you apply to a human one. Cheap hypotheses only help if the evidentiary bar holds. Otherwise you have simply industrialised the production of plausible-looking wrong answers.

Key takeaways:

Every story above is a productivity story. This one is the cost line.

Noam Segal brings the second annual tech worker sentiment survey to Lenny's Podcast - around 6,000 respondents, and year-on-year numbers that are hard to read as anything other than a warning.

Burnout rose from 44.7% to 55.7% in a single year. Optimism about roles and careers fell from 54.8% to 48.7%. The workforce now splits almost exactly in half across four archetypes: the Energized, the Conflicted, the Disoriented and the Resentful.

What drives that split is the surprise. AI-driven identity shift - the question of what your job even is now - outweighs both manager quality and company size. Those are the first two levers most leaders reach for, and this year neither was the one that mattered.

Set that against Turner's 80% SaaS cut and his headcount plan of 650 down to 400, and Singhal's ten-million-dollar offers below. The gains and the costs of this transition are landing on very different people.

Key takeaways:

The Skip runs three coaching conversations this week, and one number cuts through: "three offers go out to product executives at ten million dollars a year."

Nikhyl Singhal's framing is that executive pay has detached from the old bands entirely, mirroring what happened to AI researcher compensation. But the advice attached is more interesting than the headline. Working at the most current company beats holding the biggest title at the best-known one - proximity to where the work is actually changing compounds faster than seniority does. And on the money itself: assume pay like this comes with strings.

The case he pushes hardest on is a leader who turned down a high-growth AI opportunity over valuation concerns and then found herself choosing between safe incremental moves. His view is that she should have taken the risk - not for the upside, but for the skill-building.

Key takeaways:

A good closer, because it flips the usual framing.

Ben Hansford is a commercial director, AI filmmaker and USC professor, in conversation with Reid Hoffman and Parth Patil. He was using generative tools before DALL-E and Midjourney went mainstream - and he used them to sharpen his commercial treatments and win jobs, not to shave costs off jobs he already had.

He traces Hollywood's arc from student protests and flat industry refusal to the current scramble, as shrinking budgets and shifting audience behaviour push creators toward workflows they had previously rejected outright. He now teaches one of the first AI film courses in the United States.

In every other section this week, cheaper production shows up as displacement. Here it showed up as a pitch that beat other pitches. Same collapse in cost, opposite result, entirely down to where in the workflow it landed.

Key takeaways:

———————————————————————————

Nine conversations, one week, and a fairly consistent shape underneath them all.

The cheap thing and the scarce thing swapped places. Producing a contract, a CRM, a therapy candidate, a film treatment, a working feature - all collapsed in cost. What did not collapse: understanding a system nobody has read (Horthy), knowing which layer you are competing on (O'Driscoll), holding an evidentiary bar (Gómez-Bombarelli), and keeping half your workforce out of the Resentful bucket (Segal).

The budget did not disappear, it relocated. Curative cut SaaS 80% while its model bill grew sixfold a month. If you are planning 2027, that is the sentence to take away. Ask where your customers' spend is moving to, not just what they are cutting.

Your churn risk now sits upstream of your CRM. If a customer's internal agent can inspect your API surface and rebuild your useful 20%, that decision happens in their coding environment months before it reaches a renewal conversation. Audit yourself the way their agent would.

Put humans where benchmarks cannot see. Horthy's dark factory failed at architecture because architecture is exactly what unit tests cannot evaluate. Concentrate review at design and system boundaries instead of spreading it thinly across every generated line.

And watch who is paying. Ten-million-dollar executive offers and 55.7% burnout are the same story told from opposite ends. If your plan books the productivity gains, book the transition costs against them too.

Want more of the same? Try these five:

That’s a wrap.

As always, the journey doesn't end here!

Please share and let us know what you liked or want changing! 🚀👋

Alastair 🍽️.