Being a Unicorn is no longer good enough; you gotta think bigger

Let’s start with talent, because talent tells you where the perceived value went. The highest-agency builders, the people shipping the interesting stuff right now, will not join a 1-billion-dollar startup. A unicorn is a participation trophy. You need line of sight to 10, 20, 100 billion, or the best people would rather do their own thing. RSUs and tender offers do not move someone who can spin up a company in a weekend.

The flip side of that is the founder pool got older and wider. The 40-something second-time founder, the profile VCs used to pass on, is now the profile that wins, because agency plus cheap tooling beats a thin resume and a demo. Anyone with enough agency can go build the thing. That is the good news, and it is also the problem, because it is true for everyone at once.

The leaky RevOps bucket problem

Building is now easy, even if shipping is harder. Selling is still brutal. And churn is quietly killing the companies that figured out the first two.

You get the momentum, you land the logos, and then the customer churns 90 days later. Hard-earned revenue leaks out the bottom as fast as it comes in the top; you burn cash chasing net-new to replace what you lost, and you never reach profitability. More startups exist than ever, which means the scarce thing is not another product. It is distribution, it is trust, and it is product velocity, which is exhausting. VCs still want the 100x exits, not a tidy 10x, so the pressure to grow never lets up.

Token maxing is out, and Routing is the cool kid now

For a while, the move Big Tech wanted from its employees was token maxing, throwing the biggest frontier model at everything, driving up leaderboards, and paying for it. The unconstrained era is ending with applause from CFOs across the enterprise. The interesting question now is not “how much can we spend,” it is “what is the right model to execute this specific task.”

The pattern taking shape is that cheaper open models like the GLM family do the day-to-day work, and when they hit something hard, they make a call to a smarter frontier model, Opus or GPT 5.6 Sol, for the PhD-level help. Trained on your company data, the cheaper workers perform as well or better on the routine stuff at a fraction of the price, reportedly around 20 percent of the cost.OpenRouter sits in the middle of this and has gotten noticeably more valuable in the last three months than it looked before.

Here is the part that surprises us. Amazon Bedrock and Azure sit on catalogs of serverless, token-based models, exactly the right place to build a smart routing layer that reads the prompt and sends it to the right model, with guardrails and evaluations attached. Bedrock has the beginnings of this with Prompt Router. Neither has fully run at it. The incentive is the tell: a good router means the customer spends less per call, so the hyperscaler bills less per call. Except that is the Jevons paradox trap. Make tokens cheaper, and people do not spend less; they spread tokens across more use cases, like peanut butter, and total spend climbs. The lab that builds the best router does not lose the bill. It captures the next 10 use cases the old price tag was blocking.

The AI SDR autopsy

Nothing this year makes the “keep a human in the loop” case better than the autonomous sales development rep.

The pitch was clean: fire the SDRs, let agents prospect for pennies. The reality, based on reported figures circulating, was a glorious bloodbath. One heavily funded autonomous vendor watched most of its revenue evaporate at the 90-day mark, a reported 78 percent churn rate. Reply rates for AI-only outreach were in the low single digits, compared with a healthier human range. The cost per opportunity went the wrong direction once you accounted for the wreckage, reportedly climbing from the mid-hundreds into the low thousands.

The lesson is not “AI does not work in sales.” The lesson is architectural. You build the process around the human, so the human sits at the one decision point that requires a human, and the agent handles the pre-work and post-work. Automate the task, not the judgment. The companies that fully automate a bounded task win. The ones that get an agent 80 percent of the way through a nuanced job and call it done are the ones hemorrhaging ARR.

Salesforce quietly became the brain

Going headless, transcripts of every customer interaction, and large models may be what finally delivers the CRM brain we were promised decades ago. Pull in the real-world context, automate the data entry reps hate, feed the unstructured signals across systems- ticket data, product-led usage, churn flags, the stuff that never made it into a clean field- and the CRM stops being a database you resent and starts being the actual nervous system for the revenue org.

The AI assistant becomes the way you talk to the system, and the dashboards become live artifacts that refresh automatically. One of us already runs a fully automated weekly business review that pulls together website analytics, revenue data, Slack notes, and meeting transcripts into a single narrative with a single command. Two, the seat math changed. Jason Lemkin has talked about reducing his Salesforce seats to the three people who enter data, with everyone else on a read-only layer via the assistant. An API governance layer like MuleSoft is what lets the whole team see what one power user builds without buying everyone a full seat.

$9 billion dollars for babysitters - My kids’ futures are looking up!

If the labs believed adoption was going fine, they would not be doing this.

If investing 100% of free cash flow in CAPEX wasn’t enough, the hyperscalers said, hold my beer. Amazon invested $ 1 billion in a forward-deployed engineering org. Microsoft answered two days later with 2.5 billion dollars and 6,000 people in a new unit called Frontier. Anthropic and OpenAI stood up their own deployment vehicles earlier, reportedly around 1.5 billion and 4 billion, respectively, though the OpenAI figure has been reported as high as 10 billion depending on the source. Add it up, and you land near 9 billion dollars aimed at one thing: sitting inside customer companies and dragging AI from pilot to production.

Read the signal from the noise, not the press release. This is the realization the tech is outrunning customers’ ability to implement it. Adoption is slower than the roadmap calls for; companies cannot figure it out on their own, so the industry is back to throwing human bodies at a change-management problem. That mirrors electricity in the early 1900s, when the factories had the power and still spent 20 years learning how to rewire the work around it.

We debated whether a forward-deployed engineer is just rebranded professional services? Partly, yeah. A true FDE builds features on-site that flow back into the core product for all customers. If the work does not make it back into the product, you hired an expensive consultant to implement a system. Both can be true: some of this is real platform-building, and a lot of it is consulting with a better logo.

The narrative flipped from fear to optimism

Watch what the labs stopped saying. The early playbook leaned on fear, the sky-is-falling framing that helped justify the CapEx and the raises. The blowback showed up on schedule: data center NIMBY fights, and a workforce asking why it should adopt the thing pitched as coming for its job.

Now the message is pivoting to optimism, the “it will not take all the jobs, it will cure cancer” register, closer to the tone that would have landed better from the start. Contrast that with China’s “this is for the collective good” framing and you see two opposite adoption strategies playing out in public. The marketing shifted because fear turned out to be a bad on-ramp for a technology that you need millions of people to willingly adopt.

Vibe code has a production problem

Demos and slideware are magic. Production is a different sport.

The gap is the undifferentiated heavy lifting: the CI/CD, the databases, the identity and access wiring that turns a weekend build into something a business can run on. Until the platforms swallow that scaffolding, most vibe-coded output stays a skunkworks project, not a product.

And the claim that “vibe coding democratizes software for everyone” is too strong. A healthcare firm is elite at healthcare because that is where it makes its money, not because it is secretly a software shop. The people moving to the frontier of software development will be better at software, same as always.

The barbell: headless on one end, taste on the other

The cleanest frame we saw all week came from the GTM writers: AI is barbelling go-to-market. One end is headless. Salesforce and HubSpot are already heading there, where agents run on the platform, or the platform runs your agents, with a UI still hanging around for the humans. The other end is taste.

When execution is easy, differentiation has to move somewhere, and it moves to the thing that cannot be copy-pasted. You can already spot the AI-built sites. That is a Claude site; that is a Gemini site; the same peanut-butter schema smeared across a thousand pages unless someone fed it a real style guide and real judgment. Taste, brand, and a point of view become the moat precisely because the model hands everyone identical raw output. Teams that can encode their design and style into the system, and orchestrate on top of these headless platforms, get the edge. Everyone else ships slop at scale.

Trust is the moat AI cannot fake

Growth is now a trust problem. When execution commoditizes, value migrates to what AI cannot fake, and buyers, flooded with AI slop, notice the brands that get it. We feel it ourselves as buyers, drowning in auto-generated updates nobody asked for.

So who captures the value when building is free? Not the model; the model is becoming a commodity input everyone rents. Not the skinny wrapper, maybe a substantial harness? We know value goes to firms that own distribution, whoever owns taste, and, above all, whoever owns trust, because those are the three assets a competitor cannot spin up in a weekend and a model cannot generate on demand.

People still buy from people they know, like, and trust. That was true before any of this, and it is the one moat that gets deeper, not shallower, even as the tools get better.

Be likable. Be trustworthy. Then go sell your stuff.

I have a great idea! I’ll cut cable and sign up for Netflix and maybe Hulu. Genius, saved $100+ per month.

• People cut cable to save money, and it worked…for a while.

• Everyone unbundled. Competition thrived. Now we have YouTube TV, AppleTV, Netflix, Disney+, Amazon Prime, YouTube Red, Twitch, TikTok, and Paramount+.

• The streaming platforms got their distribution, and now they are monetizing, and we pay more than ever: internet plus a stack of subscriptions.

Well, I think we’re at the same place with SaaS. SaaS is at the same place. The cost of building software has fallen to the point where the long tail of applications has become viable for consumers and enterprises. The hardest part of running a software firm has shifted from building to distribution. Network effects and moats are getting disrupted faster. PLG is fueling record, profitable growth to $100M ARR with fewer employees than ever (Lovable hit roughly $100M ARR with about 45 people), but customer churn can be high (Easy to sign up, easy to cancel for a new app). Everyone can have their own version of an application for their phone, the web, or the enterprise.

But Quinn, CFOs have budgetary control; enterprises can build their own software to replace off-the-shelf options. That means LLM spend will go up, but SaaS will suffer.

I agree. Some legacy SaaS and consulting firms will not make the transformation the same way legacy software firms failed to transition their business model to subscription SaaS. But the two forces I see challenging the “SaaSpocolapse” narrative are parity and functional expertise.

AI software becomes a productivity enhancer and a tax as every firm reaches adoption parity.

Every company will buy the leading SaaS and AI-native offerings, not because they want to, but because they have to. The productivity gains are a forcing function for executive leadership. Speed is the new exhausting moat, and LLMs become both a productivity enhancer and an enterprise tax. The long tail of profitable software means more niche offerings for every problem enterprises, departments, and SMBs face. The counterargument is cognitive load: managing a broad portfolio of apps pushes the dynamic back toward consolidation and bundling. but there is a lot of nuance in that statement, as each department can have its own set of AI software tuned to its exact problems. But that pivot is weaker than it looks, because each department can run its own set of AI software tuned to its exact problems.

Expertise is an accumulation of acquired and applied knowledge.

Retailers, manufacturers, financial institutions, and healthcare firms can build software faster than ever before. If you compare that to the last ten years of SaaS feature releases, you would be right to assume they will benefit from doing just that, building their own software. The assumption we often make when predicting the future is that everything else will remain the same. That is, existing SaaS firms will not take the same technology and accelerate their development and innovation at a multiple or exponential rate over firms whose core business is not developing software.

I do not want to leave you with the impression that enterprises and employees will revert to the old paradigm of outsourcing technical innovation to SaaS firms. Far from it. I believe their innovation will be incremental to existing employee tasks and will incorporate the firm’s proprietary data that is not accessible to third parties. That will force software solutions to be built that solve harder, tertiary problems related to the firm’s core competency. This evolving frontier will continue to drive innovation, churn among SaaS customers, and productivity across the economy.

It is an exciting and exhausting paradigm we are walking into. It’s a version of the “productivity paradox” documented by Nicholas Carr’s 2003 Harvard Business Review essay “IT Doesn’t Matter.”

So the question is not whether software spend goes up. It does. The question is who captures the value creation: the model providers, the AI-native SaaS firms, or the enterprises that turn their own proprietary data into software no vendor can sell them. My bet is on all three, but the majority goes to the firms that execute the right strategy the fastest. And the Net result: everyone will spend more on software, not less, in the years to come.

It’s a good time to sell; it’s a hard time to be in customer success.

We made it to another Friday, and the whiplash the past week was something else. On one side of the ledger, the frontier kept sprinting: a new class-leading model shipped, three mega-IPOs and mega-rounds soaked up a casual quarter-trillion in capital, and somebody is now building an “artificial general engineer” with enough money to buy a small country. On the other side, the people who sign the checks finally showed up to the party. We talked to a CFO this week who said the quiet part out loud: “We have budgets now.” Four words. They land like a brick.

That is the tension we want to sit in this issue. The frontier goes up and the budget comes down, and everyone selling AI is now stuck in the gap between the two. Here is what is happening in that gap, and who walks away with the margin.

Subscribe now

The retention policy nobody read until it mattered

Fable 5 shipped this week, the class model out of the Project Glasswing work, and the bench scores are strong if benchmarks are your thing. But the fine print is doing a lot of work. It is gated for legal, finance, and similar regulated workflows, and a chunk of the security tooling is walled off. The detail that moved customers: the zero data retention carve-out went away, and prompts are now held for 30 days.

For anyone operating models on behalf of customers, that is not a footnote, that is a renegotiation. If you are the operator sending those prompts upstream to the lab, your customer’s zero-retention promise just developed a 30-day hole, and they are the ones who have to explain it to their own security team. We are watching people scramble to sort this out in real time. The lesson that keeps repeating in this market: the model spec is the easy part. The data-handling terms underneath it decide whether the enterprise deal lives or dies.

The cognitive dissonance at the center of the AI economy

The companies whose entire pitch is “we will automate everything” have more open go-to-market roles than any other category. The everything-automation machine needs an army of humans to sell it.

It is easy to call that hypocrisy. We think it is something more interesting. When the cost to build collapses to roughly zero, when everyone can ship the same feature in the same week, the scarce thing is no longer the product. The scarce thing is getting a human to notice, trust, and buy. Distribution becomes the moat. The harness becomes the moat. And distribution, for now, is still made of people picking up the phone.

Look at the role everyone is suddenly hiring for: the GTM engineer. Clay gets the credit for naming it, and the shape of it is a single seller who collapses three old jobs into one. Prospector, closer, and technical pre-sales, all in one person, armed with automations. The promise is leverage: take a rep who used to need a development rep and a solutions architect riding shotgun, and hand them a territory of 1,500-plus accounts they can work.

Two things can be true here. The model is real, and it is not universal. It is a slam dunk when the company sells the exact tool the seller uses to sell. Clay reps prospect with Clay, so they can show a customer precisely how to run the play, and the product feedback loop is tight. That is dogfooding as a sales motion. But you do not spin up a petabyte data cluster to prospect, so the rep selling heavy infrastructure cannot dogfood their own product the same way. The payoff is highest in greenfield, where one person works 50 to 500 accounts and “enrich these 60 logos, find everyone in finance and procurement, draft the outreach, send the invites” is a genuine superpower. In a two-account strategic patch, the same stack is mostly noise. You already know those buyers better than any enrichment tool ever will.

The generalist is back, and the bottleneck has moved

Satya framed a related shift around Microsoft Build: the return of the generalist. We spent a century turning the village blacksmith and baker into assembly-line specialists. AI is swinging the pendulum back. One person now does web dev, hosting, sales, and pre-sales in an afternoon.

We saw it live this week. Take a seven-page proposal, the kind nobody reads because it is all SLAs and compliance boilerplate, and turn it into an interactive site with revenue sliders, so the customer can model their own pricing and watch the numbers populate. Two weeks of designer-and-developer time, done before lunch. Does a slick interactive proposal pull a deal over the line? Maybe not. But every time you lower the customer’s cognitive cost of saying yes, you are winning at the margin, and right now it separates you from the rep still attaching a PDF.

Here is the catch we keep running into, and it is the real story. Building software is become cheap. Shipping is not. The interactive proposal sat finished while we hunted for an internal hosting account and a domain to put it on. The bottleneck is no longer creation. It is the last mile: deployment, permissions, hosting, the boring plumbing that turns a cool artifact into something a customer can open. The teams that win the next 18 months are the ones who solve that last mile internally. Shopify reportedly built exactly this: an internal one-click deployment surface where any employee can drop in an AI-built app or demo, and it publishes, securely, behind the company walls. That internal deploy button is quietly one of the highest-leverage things a company can build right now.

Personalization at scale is the new table stakes, and the floor keeps rising

McKinsey’s latest B2B Pulse surveyed roughly 4,000 decision-makers across 13 countries, and the headline is blunt. Omnichannel, e-commerce, digital tools: all table stakes now, none of them a differentiator. What separates leaders from laggards is one-to-one personalization, real generative-AI use, and tight account-based-marketing governance as a control layer.

The number that matters: market leaders are about 4 times more likely than laggards to run true one-to-one personalization, 20 percent versus 5 percent. Personalization used to be the swap-in variable fields in Salesforce, and it never scaled past that. AI is the first thing that lets you do individual outreach across a real account list without the economics falling apart.

But scale cuts both ways. The same engine that personalizes 500 messages can scale a single mistake to 500 inboxes. Wrong product, wrong price, a typo that references the wrong company, now shipped at volume. The auditing layer is not optional; it is the whole job. This is why “send one more message” no longer fixes anything. When everyone can generate infinite outreach, volume is noise. The differentiator is knowing something true about the buyer and adding a human touch on top. Which loops right back to the GTM engineer: this entire motion is predicated on pulling first-party, second-party, and third-party signals into one enriched view, and most teams still cannot connect those pipes.

The 2x2 that explains who keeps the margin

The cleanest mental model we found this week is a 2x2, the framing traces to Sarah Guo. The best line in it: what you can measure, you can train against, and anything measurable slides toward commodity. What survives is the value a model cannot reach, the correctness that exists only inside one firm’s private data.

Lay it out. One axis is task difficulty, saturated versus frontier. The other is the data, public and freely verifiable, versus private with no public leaderboard.

• Frontier plus public data: the labs win, every time. Coding benchmarks live here. The best model takes it.

• Saturated plus public data: pure commodity. “Summarize this email.” Cheapest, fastest model wins, and the value rounds to zero.

• Saturated plus private data: the wrappers. Harnesses around a model, living on proprietary data that the labs cannot see. This is the contested middle, because the labs are marching straight into it, absorbing legal, finance, and security, while the wrappers race to stay ahead on domain data that the labs still do not have.

• Frontier plus private data: the promised land. Your company’s data, or the pooled data of all your customers, at a scale no one else can touch, combined with frontier capability. Nobody can commodity-price you out of this corner.

Now the part that pays. Harvey and Fireworks published a legal-AI result that makes the margin story concrete. A hybrid harness with an open-source GLM 5.1 worker, calling Claude Opus 4.7 as an advisor only when it helped, hit 18 of 100 full-rubric passes at 368 dollars total. Standalone Opus scored 14 of 100 at 954 dollars. Better quality, roughly 39 percent of the cost, and the worker pinged the frontier model just 0.83 times per task on average.

The outcome is not the point. The margin is the point. If the lab attacks a legal task directly, it charges the customer and keeps the spread. If the wrapper performs the same task at a fraction of the inference cost through smart routing, it can price below the lab, still save the customer money, and pocket the difference. That is how AI application gross margins climbed from the low 40s to the high 50s, even as raw inference costs rose. The harness captures the value the model leaves on the table.

The pricing wheel never stops turning

We will leave you on pricing, because it ties the whole issue together. The durable take, echoing comments from Satya Nadella: per-user pricing is an artifact of buyers needing budget certainty, and it survives even as consumption metering slides underneath it. Subscriptions bundle usage into per-seat stacks, with consumption below; that is the adjustment GitHub Copilot made after real usage blew past what a flat seat assumed.

Watch the cycle. 1/ Per-seat could not hold once intensity exploded, so vendors moved to 2/ usage. Usage, credits, and seats became the 3/ hybrid. Next comes 4/ outcome pricing: charging for the result rather than the tokens. And the moment outcomes get predictable, buyers will want their margin back, ask for a discount, and push the whole thing back toward fixed budgets and seats. Then we’re back to per-seat pricing.

It is a wheel, not a ladder. That sounds exhausting, and it is, but it is also the most bullish thing in this issue. Humans are never satisfied. The second a cost becomes predictable, we want it cheaper, and we invent something new to want. Somebody has to go figure out what we want next and build it. That churn is the job. As long as it keeps turning, there is work in the gap between the frontier and the budget.

So: the frontier goes up, the budget comes down, and the whole game is who captures the value in the middle. Our bet stays the same. Build relentlessly, but spend at least as much energy on distribution and the last mile, because that is where the margin lives. Just go build. Then go figure out how to get it in front of someone.

See you next week.

Fringe Lines covers the business layer of the AI industry for builders and GTM operators. Forwarded this? Subscribe at newsletter.fringelines.io.

The features buyers are evaluating have shifted from the model to the runtime: the harness, routing layer, eval system, and distribution. The weights are getting cheaper and more interchangeable by the month. Everything wrapped around them is where the moats are emerging and where value is captured.

That’s the throughline tying together a busy week: Microsoft shipping seven of its own frontier models and naming a rival framework more than two dozen times, Anthropic winding down the all-you-can-eat token era, Salesforce posting triple-digit growth on an agent product, and a 95-minute film that cost half a million dollars to generate. Different stories, same lesson. The model is the electricity. Nobody gets rich selling electricity.

Subscribe now

1. Microsoft’s new AI strategy isn’t OpenAI

At Build, Microsoft did two things at once. It unveiled seven in-house models under the MAI family, including its first homegrown reasoning model, and it built its always-on agent, Scout, on top of OpenClaw, the open framework Satya Nadella had called a “virus” three months earlier. Read the keynote and the subtext is loud: Nadella wants every company to graduate from consuming frontier models to building with them, and Microsoft is going first. The pitch is “your own model, in your environment, on your data, under your control.”

What’s interesting to us is who got name-checked and who didn’t. The open framework came up constantly. The longtime partner barely registered. When the most strategic company in software starts treating your model as one swappable input among several, that tells you where it thinks the durable value sits. Hint: not in the weights.

This is the part people keep missing. We’re now in phase three. Phase one was the model. Phase two was the harness, code generation being the first real one. Phase three is distribution: how fast you get an audience and scale, and you have to do it while you’re still building the product, not after. Product-led growth still matters. It just isn’t the moat anymore.

2. Anthropic ends the all-you-can-eat token buffet

The Golden Corral era is closing. Anthropic is moving its headless and agent-SDK usage off the bundled plans and onto separate, metered, API-rate credit pools. If you’ve been running anything autonomous through a third-party harness or the agent SDK and assuming it draws from your flat subscription, that assumption just expired.

Here’s the thing nobody wants to say out loud: the subsidies were never as generous as the internet claimed, and they’re getting thinner. There’s no magic margin to give away. GPUs are expensive, the labs are lighting money on fire, and a flat-fee buffet only works until enough people eat. Interactive use self-limits because a human is sitting there watching. Unattended agent jobs do not, which is exactly why they got their own meter.

We’ve watched this movie before. It’s cloud, beat for beat. The buzz, the “you can do anything” phase, then the bill arrives and suddenly you’re standing up a FinOps org, buying savings plans, hunting zombie VMs, and feeding everything into a cost dashboard. The technology isn’t bad. It just needs a whole second layer of systems and guardrails to run it sanely. We’re about to get the AI version of Cost Explorer, and somebody is going to build a very good business selling it.

3. The token-maxxing reckoning is here

For months the feeds have been wall-to-wall with “how to spend fewer tokens.” That conversation just stopped being theoretical. When Uber blew through its entire 2026 AI budget in four months, the response was a $1,500 monthly cap per engineer, per tool, tracked separately for Claude Code and Cursor. That’s not a ban. It’s budget governance arriving the way it always does, late and all at once.

We think the caps are healthy. You always gated your expensive infrastructure. You’d never hand an intern unlimited access to your Spark clusters to run Hadoop jobs overnight, because the overruns were real. AI got a temporary FOMO exemption from that discipline, and it was never going to last.

The fix isn’t austerity, it’s routing. Stop sending every request to the frontier. Set your default to a mid-tier model, send maybe 20% to the top, and use deterministic workflows where a workflow will do. If you’re asking how to do something in Linux, that’s a Google search, not an Opus call. The emerging pattern we keep hearing about: a cheaper open-source worker model does the bulk, then calls a frontier model as an advisor or reviewer on a tool call, rather than letting the expensive model run unsupervised. Harvey runs open-source workers and calls a top model in as a tool. Replit and Lovable, both brutally cost-sensitive, lean on the same trick: generate with a cheaper model, then point a second model at the output to check and fix it.

But here’s our worry, and it’s the cloud parallel again. The routers have intelligence, sure. How do you know the router is optimizing for your business and not just for cost? To trust it you need an eval system, people who understand the eval system, and people who understand what good output even looks like so they can confirm the evals are tuned. Congratulations: we just reinvented model drift from traditional ML inference. We’re stuck in the same loop, one layer up.

4. Salesforce proves the headless agent business is real

Agentforce crossed $1.2 billion in ARR, up 205% year over year, the first time it cleared a billion. Stack that against the combined ARR of basically every AI-native startup and it’s a genuinely impressive number from a company whose core is growing single digits. Salesforce is pivoting hard into agents, and Service Cloud, Zendesk, HubSpot, and everyone else are right behind it.

So here’s the question that matters for GTM: are these agent-focused, headless platforms going to grow single digits, double digits, or triple digits? Salesforce just answered with triple. The future is headless. You’ll still have a UI and seats for humans, but you now have to assume agents are coming through the API and through MCP, in volume.

Now sit with the scoping conversation that creates. How many agents do you have? How many need access to Salesforce and your database? What’s the concurrent-call load? Ask most orgs today and you’ll get a blank stare, because the systems probably aren’t even throwing off the telemetry you’d need to answer. We talk endlessly about observability, and we still don’t know if these stacks emit the signal required to make an informed decision. We are not there yet. That gap is a business waiting to be built.

5. A new 100% AI film is a $400,000 argument for the harness

A 95-minute fully AI-generated film called Hell Grind reportedly cost about $500,000 to make, of which roughly $400,000 was compute. (It was shown at an industry event in Cannes, not the official festival program, despite some breathless early coverage.) The number we can’t stop thinking about isn’t the budget. It’s that the average prompt ran around 3,000 words, and the team needed real cinematography knowledge to write them. Shot composition, sequencing, why you can’t run two close-ups back to back.

That is the whole thesis in one production. The startup behind it doesn’t even make the video model; it uses Google’s Veo 3 and builds the layer that keeps lighting and characters consistent across a feature. The model was the commodity input. The harness, plus the human who knew what to ask for, was the product. You can’t hand a toddler Claude Code and get a business, and you can’t hand us a camera rig and get a blockbuster. The orchestrator, the subject-matter expert, is still load-bearing.

If AI is a utility, how fun is selling electricity?

Here’s the comparison everyone keeps circling back to. LLMs are a general-purpose technology, like electricity. Fundamentally transformative, genuinely life-changing, but how many of us dream of quitting our jobs to go sell electricity? It isn’t fun, and it isn’t differentiated. So the question for every operator this year is: are you selling electricity, or are you the electrician who wires the building, the one who brings domain knowledge to solve a specific problem in a way that’s never been done before?

Because the macro hasn’t moved to save you. We did not get twice as many buyers or twice the budget this year. GDP grows around 2%, which means next year’s pie is roughly 2% bigger, full stop. So you either take revenue from someone else, find real productivity, or figure out how to stand out from the noise. The building of software got cheap. Distribution didn’t. The harness is the moat, and the value goes to whoever owns the runtime and knows what to do with it.

Keep your head down and keep building. There are still real, unsexy problems to solve in this supposedly pre-AGI world, and the people who solve them are the ones who’ll capture the value. We’ll see you next week.

Subscribe now

Fringe Lines covers the business layer of the AI industry for builders and GTM operators. Reply and tell us where you think the value is getting captured.

Here is our read on the topic that caught our eye this week across Sales compensation, usage vs. contract bookings, and Tokenmaxxing. But first, the headlines…

Fringe Lines is a reader-supported read on AI, builder tooling, and go-to-market. If a friend forwarded this, subscribe.

Subscribe now

1. The comp bubble at the top, and what it tells you about the sustainability of AI-first companies

Start with the number that broke our brains this week. Top-tier CROs AI startups are reportedly fielding packages up to $100 million. Bubble anyone? The entire premise of the AI sales era is that the motion is automating itself: AI SDRs, automated lead enrichment, self-serve onboarding, doing the work a body used to do. So why is the person at the top of that org worth nine figures?

The answer is the tension we kept circling back to. We heard this week that a lot of these AI companies in exponential growth are running group quotas, not individual ones. No personal number. There are decent arguments for group quotas. How do you set quotas for once-in-a-generation growth percentages? I can tell you, finance doesn’t want to pay the accelerators, so they default to equity. If the company wins, you’ll overachieve in the IPO.

That works right up until it doesn’t. Our concern, and we are not alone here, is the sales-socialism challenge. If nobody carries an individual number, you slowly accumulate quiet underperformers who have never had to win a deal on instinct. A real sales culture culls the bottom 10% on a rolling basis based on performance, so that overachievement, president’s club, and grinding mean something. Strip out accountability, and the whole thing looks great while demand is a tailwind, but the faster revenue and production adoption ramps, the faster it can drop if you don’t have strong moats to maximize switching costs (network effects, distribution, scale economics).

Compensation aside, one if the interesting takeaways from 2026 is counterintuitive, but an optimistic one. AI will not replace the rep. The flood of AI-generated outbound made cold digital noise nearly worthless, which raised the premium on the human who makes the phone call, leaves the voice mail, sends the text, and shows up in the room. Human interaction is the differentiator, and you can see this in the percentage of sales job openings among total openings at the fastest-growing AI companies.

Chart courtesy of GTMnow.

2. Contracts are friction. They are also the only thing stopping a one-click exit.

The comp story rests on a deeper one about contracts, and this is where it gets uncomfortable for fast-growing AI PLG companies that prioritize growth over commitments. Inserting commitments and contracts into your sales cycle adds friction on the front end, elongating the sales cycle and therefore slowing growth. It also acts as friction on the backend, helping reduce churn. You need a balance of commitments with Account Teams, also held responsible for usage. Otherwise, you risk shelfware or commitment shortfalls that churn because the economics are so bad.

The former CRO who took Snowflake from zero to $4 billion framed it cleanly: contracts act as friction against churn. They buy you time. A customer under contract cannot quietly decide on a Tuesday to migrate off and never speak to you again. They have to come to the table, just as a data center lease forces a conversation. Month-to-month is more expensive precisely because optionality has a price.

Now layer in what we heard recently at SaaStr’s conference, more than half of Anthropic’s enterprise customers are self-serve onboarding. I believe they did this out of necessity, can’t scale hiring and onboarding fast enough, and because they weren’t willing to lower their bar for talent. The challenge is that Anthropic has OpenAI chasing after more Enterprise use cases with its Codex offering, desktop app, and availability on Amazon Bedrock. If half your enterprise base is not talking to a rep, then half of it is almost certainly not signing anything either. Which means the thing that used to protect you on the way down has been deliberately removed in the name of frictionless growth on the way up.

Founders will tell you that contracts add friction to the acquisition process, and they are right. It is a genuine double-edged sword: it helps you climb, and it hurts you when you fall. But the version of this that should worry an AI seller is the customer who says, “We love your models. We put OpenRouter in front of everything, pay a small premium, and cut our costs by 40% by routing the easy work to cheaper models. Just keep making great models, we will consume them our way.” What is your rep supposed to say to that? “Do you have a contract?” “No, we’re good.”

There are options for the Labs. You can gate your newest models. You can stop subsidizing third-party harnesses, which is exactly why Anthropic pulled third-party harness access off its most subsidized consumer plan: why pay to drive adoption of someone else’s wrapper. The harness is becoming a real moat. But every one of those answers points in the same direction. The percentage of enterprise customers you actually talk to has to go up from 50% as you scale your Enterprise Sales team. The other challenge with commitments, customers expect discounts in exchange for the lock-in of usage. That is another challenge for Labs as they strive for profitability ahead of an IPO. A 5% discount doesn’t sound like much unless your EBITDA is 5% as well.

3. Tokenmaxxing met the finance department in 2026, and your CFO is not impressed

AI optimism above runs straight into a wall of worry this week, the AI bill is coming due.

A stat to carry into Monday, courtesy of CFO Pilot citing Ramp’s spend data, is that the average business is now spending 13x more on AI tokens than it did in January 2025. Less than eighteen months, basically zero to the largest vendor line on the sheet. Uber’s COO publicly said the company burned through its entire 2026 AI budget in four months and that the link to shipped consumer features “is not there yet.” Amazon reportedly scrapped an internal AI usage leaderboard, with leadership telling staff not to use AI just for its own sake. Microsoft reportedly canceled Claude Code subscriptions for several product divisions. Fortune declared the era of measuring AI value by tokens consumed officially over.

This is Goodhart’s law playing out in real time. When a metric becomes the target, it stops being a good metric. Hand an org a token leaderboard, and people will run it just to run it. The fix is not spending less; it is spending like an adult. CFO Pilot’s line is the keeper: “You don’t use a heart surgeon to take your temperature.” A $100K-per-month Opus bill drops by roughly 48%, which is $576K of savings on an original 1.2M run rate, by routing easy work to cheaper model tiers. Set the org default model to the middle tier, and 90% of users just stay there, which is up to 40% in savings before you do anything clever. Add per-user spend caps, prompt caching for repeated contexts at roughly 90% off the cached portion, and the Batch API at 50% off for anything that does not need to happen in real time.

Finance is getting its arms around tokenmaxxing in 2026. 2027 is going to look very different.

4. How the Hyperscaler Capex gives insight into differentiation

The framing that lit us up this week, via SemiAnalysis, is who actually makes money on the economics of tokens. The first tenet: tokens carry more profit for a hyperscaler than raw infrastructure-as-a-service, because if you are just renting GPUs, you are largely subsidizing the chip vendor. Tokens-as-a-service is a fundamentally better business model than the IaaS contract it sits atop.

Then it becomes a question of whose silicon is being used in each cloud. Amazon has Graviton for general compute and Trainium for training and inference, and a large share of its managed inference reportedly runs on Trainium, which means it is dodging a chunk of the NVIDIA tax. Google has TPUs doing a similar job. Microsoft is still largely NVIDIA, AMD, and Intel while it builds its own. The Neoclouds, CoreWeave, and Oracle tiers are essentially NVIDIA resellers. Advantage: the two players who built their own silicon.

Now let’s pivot to capacity to handle customer GenAI workloads, which includes energy, data centers, compute, and silicon. Amazon and Google moved first to lock in energy commitments, so they had the capacity to open data centers when demand showed up.

Next, let’s look at how each company is meeting their customer demand for capacity (Internal and external).) Microsoft has to host its own Copilot, GitHub Copilot, and OpenAI inference. Google has to power its own search, its Frontier training, its Gemini Workspace assistant, and Claude models on top of GCP. Amazon, by comparison, has more of its built capacity free to sell, and in a world where demand still outruns supply, that surplus converts straight into margin. Net: AWS operating margins expanded meaningfully quarter over quarter, while peers went flat to down. Hypothesis? Amazon’s vertically integrated stack, combined with Amazon Bedrock's hosted model revenue charged to end customers rather than internal business units, is driving token profitability.

We cannot independently verify the full profitability ranking, and the most interesting numbers here are extrapolations rather than disclosed figures, so take them with a grain of salt. But the strategic shape is sound, and it reframes the buying conversation. It is not just “Claude or OpenAI.” It is “which cloud are you buying through, and what does that do to your effective unit economics?” Most buyers are not asking that question yet. They should be. Check out SemiAnalysis’s article for a deep dive into the token economics.

Where this leaves us

The thread connecting all of this week’s topics is a pendulum. We are at peak GenAI hype (tokenmaxxing) on the Gartner hype curve, the part where horizontal AI will eat all Software and jobs globally, and we think the swing back toward reality is starting. That is healthy and refreshing. The cortisol of the past six months is starting to take its toll. The reason both major labs entered into joint ventures with consulting firms is that GenAI/LLM adoption and ROI are genuinely hard. You do not peanut-butter-smear AI tokens onto a problem and magically get your ideal outcome.

The analogy we ( read everyone) keep coming back to is electricity. If large language models are the general-purpose technology, the raw current, then what we are still missing is the appliances, the specific products that convert that current into productivity. Right now, we are at the light-bulb stage: personal output per person is up, but companies have not figured out how to capture that gain at the team, function, or firm level. The honest possibility is that there is a real ROI already here and simply not surfacing, because the person who saved two hours drafting feedback with Claude is reinvesting that into their next AI project or taking a longer lunch. People could be quietly getting more done and not giving the time back to the employer. I think a more nuanced and balanced attestation is that more people can do more tasks outside their core domain or skill set. This won’t immediately lead to a productivity boon, but as employees skill up on a broader set of tasks, the productivity will trickle through the firms.

If we are right, the revenue curve flattens as customers optimize their spend, then turns back up sharply once verifiable ROI becomes clear and the rush to scale use cases with provable ROI begins. The open question, the one that decides the winners, is whether the labs can become cash-flow positive fast enough to still be standing when that second wave arrives. With the amount of capital being pulled out of the market this year from the IPOs, it’s an open question.

Come back for more sauciness next week.

Subscribe now

Fringe Lines is a reader-supported read on AI, builder tooling, and go-to-market. If a friend forwarded this, subscribe.

One of us opened this week’s show with the thing we keep circling back to: the gulf between people who are actually AI-pilled (using these tools every day, shipping with them, trusting them with real work) and everyone else is widening. Every week. It’s a little terrifying, because the tools make the pilled group powerful in a way that’s genuinely hard to articulate if you’re on the other side of the canyon.

The other one of us put it differently: you have your normal work, and then you have this second firehose pointed at your face all day long, and you’re just trying to get a sip down without drowning. The water isn’t going in your mouth. That’s the whole AI-builder mood in April 2026, summed up.

This week, three things happened that all tell the same story from different angles. Pull up a chair.

1. Anthropic walked into Figma’s boardroom, then walked out with the product

Here’s the timeline, because it’s kind of incredible.

Mike Krieger is Anthropic’s CPO. He also sat on Figma’s board. Last week, on Wednesday or Thursday, he stepped down. Submitted his resignation. A day or two later, Anthropic shipped Claude Design, which is, let’s not be cute about it, a Figma competitor.

That’s a fucked up little conflict of interest. You can file it next to Eric Schmidt sitting on Apple’s board right up until the Android launch. Design cycles take longer than app cycles, so the CPO had to know what was coming. The fact that it landed within 48 hours of the resignation is not an accident.

Whether you think it’s a dirty move or just the market working at its current temperature, the real signal is velocity. Anthropic is now shipping a product a day. If you’re a frontline seller anywhere in the tech stack and your customer says “what about this new Claude thing?”, there’s a real chance you hadn’t heard of it when you woke up that morning.

2. Uber’s CTO blew his entire 2026 AI budget by April

This one is not hypothetical anymore. Uber’s CTO, Praveen Neppalli Naga, admitted it publicly: Uber’s full-year AI budget is gone before summer. Claude Code burned through it.

The math, per the reporting: Uber gave 5,000 engineers Claude Code access in December 2025. By February 2026, usage had nearly doubled. Four months after rollout, the annual AI budget was completely gone. Currently, 95% of Uber engineers use AI coding tools every month, 70% of committed code originates from AI tools, and 11% of live backend updates are AI-written. That’s 1,800 AI-generated code changes per week flowing into production.

The CTO’s quote is the tell: “back to the drawing board.” A company with a $3.4B R&D line and dedicated budget planning teams missed the trajectory that badly. If Uber missed it, your CFO missed it too.

And this is where token anxiety stops being a meme and becomes a line item. The 11 Labs CRO floated a take that stuck with us: if everyone throws a hundred-x (or a thousand-x) more agents at the same buyers who didn’t want to be talked to in the first place, that’s not a recipe for success. It’s a recipe for margin erosion dressed up as productivity.

So here’s the question nobody’s answering out loud: where’s the incremental revenue coming from? Engineering budgets are up. Productivity is up. Headcount is mostly still there. CFOs are running the same spreadsheet and arriving at the same place: you’re speeding up, but the top line has to catch up or something gives.

Our bet, call it our pessimistic 2027 prediction, is that the labs pull a Coke-and-Pepsi. First you open the system, get everybody on the platform, build lock-in. Then you close it and raise prices. The trigger will be layoffs: once enterprises free up budget by letting people go, the labs will know there’s newfound room to charge, and they’ll take it. Layoff day + price hike = revenue shoots through the roof.

3. The GTM teams already living in 2027

If you want to see what the other side of the gulf looks like, Brendan Short at The Signal put on a live event this week with six GTM operators showing off what they’ve actually built with MCPs + Claude. It’s the best snapshot we’ve seen of what “AI-pilled GTM” actually means in practice. Rough inventory:

Team What they built Why it matters Eric Fitz (SMB AE, Zendesk) Sumble MCP + Claude Cowork → identifies call-center prospects in his 1,800-account territory, drafts tech-stack-referenced emails 99.8% deliverability on ~5,000 emails, new edge on a territory that had been reused for 5 years Josh Nelson (RevOps, Nooks) DIY data warehouse on Render + Redash + custom MCP → Claude queries HubSpot, Mixpanel, Nooks, CPQ in plain English What would have taken a DBA two years, shipped over a hackathon weekend Vinayak Mehta (SWE, Clarify) Clarify MCP → iMessage screenshots become CRM entries, LinkedIn ad engagers flow into outbound, contracts auto-update deal fields The rep never leaves the Claude window Anis Bennaceur (CEO, Attention) Call-recording MCP + Clay → auto-clusters persona pockets with Claude-generated names (”The Frustrated Veteran”) Daily scheduled runs; every customer call compounds into better outbound Sam Gong (SVP Mktg, WorkSpan) Octave MCP → 60% of their shared context layer auto-populated from SKO transcripts in one hour Killer quote: “I would not have bought this tool without the MCP.” Jiquan Ngiam (CEO, MintMCP) 25 Slack agents, GitHub as long-term memory, Attio MCP read-only for governance 1-to-5 human-to-agent ratio across the whole company

The thing you can’t miss in this group: the MCP is doing the work of the moat. Six months ago a good product was the moat. Now a good product with a good MCP is the moat, because the product is what gets plugged into Claude and the MCP is what makes it actually useful inside the conversation your rep is already having. If your vendor doesn’t ship one, they’re about to get skipped.

4. Meanwhile, your kids already picked Gemini

One of our kids is seven and currently pulling us toward Gemini: “look it up on Gemini, let me design my images on Gemini.” At school, it’s all Chromebooks, Google Workspace, and Adobe Express. My daughter sat down with Google Stitch last night and was immediately generating app prototypes without instruction.

We keep forgetting this: the sleeper in the model race isn’t Claude or OpenAI. It’s Google, because Google already owns the desks, the keyboards, and the browsers that eleven-year-olds are sitting at every morning. Gemini hit 750M users, 3.1 Pro shipped globally, they dropped a Mac app, and Antigravity (their IDE) now bundles Claude models alongside Gemini and Nano, with the kind of Chrome integration you only get when you own the operating surface. A Gemini subscription gets you Opus 4.6 in the IDE. That’s a weird sentence to write, but here we are.

Distribution eats model quality for breakfast. Always has. Always will.

5. What we’re trying this weekend

Builder energy is peaking and so are the subscription bills. Here’s the shortlist of what’s actually earning its $10-to-$50/month right now, and what you can ship without paying a dime.

Google Stitch: free UI/UX prototyping on a Google Labs URL. Prompt in an app idea, get iPhone mockups, export to AI Studio. Not great portability to Figma, but incredible for ideation if you’re not a designer. Run it with a seven-year-old. Trust us.

shadcn + Claude Code: you do not need to pay $300 for Tailark. Point Claude at the shadcn library, describe the aesthetic, let it cook. Our recent WordPress-through-Claude-Code refresh was maybe an hour of “accept, accept, accept.” Dark theme and light theme, pattern-matched to a reference site we liked.

Antigravity (Google): if you’re bouncing between Chrome, logs, and an API, the browser integration for troubleshooting is genuinely the best we’ve used. Built on VS Code. Agent sessions organized by project folder, same as Claude’s new Code tab.

Framer: $10/mo including domain. For anyone who is tired of the WordPress/Wix/LightSail triangle and just wants a static site up. One of us is publishing a Claude-built fantasy-football-101 interactive guide this weekend purely because the niche is wide open and the tooling makes “publish” a one-evening affair. (Monetization play: affiliate links to DraftKings. The resource page writes itself.)

Paper Design: a likely company-of-one, $10/month, nails UI/CSS styling in a way Figma has to work at. Probably profitable from day one; watch it.

The pattern is the same across all of them: it’s a weekend to ship now. If your moat was “it takes three months to build this ad-generation pipeline,” that moat is gone. Opus 5.1 will close it further next quarter.

The thing we keep landing on

Ben Horowitz was everywhere this week describing the same thing we’ve been circling on the pod: “If you keep looking at it like the old world, and it’s got completely different laws of physics, you are definitely going to die.” He was speaking to founders about the anxiety of not moving fast enough.

But there’s a second anxiety he nailed that’s harder to talk about: workers are afraid of becoming irrelevant, and more than 54% of workers bypassed their company’s AI tools in the past 30 days and completed the work manually instead; another 33% haven’t used AI at all. Combined, roughly eight in ten enterprise workers are either avoiding or actively rejecting the technology their employers are spending record sums to deploy.

Two anxieties. Same machine. No one bridging them.

That’s the gulf. It’s not going to close on its own, and nothing in the current release cadence is slowing down to let anyone catch up. You either build something this weekend, or you watch your seven-year-old do it for you.

See you next week. Stay hydrated. The firehose isn’t stopping.

Quinn & Doom

Referenced this week:

• Brendan Short, The Signal Club: 6 Ways GTM Teams Are Using MCPs and Claude Today

• Nick Lichtenberg, Fortune: a16z’s Ben Horowitz sees ‘AI anxiety’ consuming Silicon Valley founders

• Forbes / Josipa Majic: Ben Horowitz says America must rebuild and AI hits a bottleneck everywhere

• Callum Williams (@econcallum) on X: original post

• @mstockton on X: Uber CTO on blown AI budget

Let that sink in: a company that was considered the “laggard” eighteen months ago is now arguably the hottest enterprise AI company on the planet.

Claude Mythos & Project Glasswing

Anthropic also debuted Claude Mythos — their most capable model yet — alongside Project Glasswing, a security-focused initiative that represents a step function change in cybersecurity capabilities. Mythos is reportedly so effective at discovering zero-day exploits that Anthropic is giving preview access to large tech companies and roughly 40 additional organizations building software and maintaining open source projects, so they can patch vulnerabilities before the model is widely available.

Our hot take: It wouldn’t surprise us if running a Mythos-grade security scan becomes a compliance requirement within the next 18 months. And if it does, model availability will likely become gated and premium-priced — we haven’t seen the labs restrict access like traditional SaaS feature tiers yet, but this feels like the kind of capability that changes that.

OpenAI’s Enterprise Problem (And Why It Might Not Be Their Fault)

Here’s the thesis Quinn has been building: OpenAI may be contractually handcuffed in the enterprise.

When OpenAI signed its deal with Microsoft several years ago, Microsoft became the only hyperscaler that could host OpenAI models. Fast forward to today and the competitive landscape has completely inverted. Anthropic signed deals with Amazon, and now every major cloud has Claude in its lineup. AWS sellers push Claude on Bedrock. GCP reps sell Claude on Vertex. Even Microsoft sellers can offer Claude on Azure Foundry.

Meanwhile, OpenAI’s only enterprise advocates are Microsoft reps — who are incentivized to sell a buffet of models, not just OpenAI. The result? Tens of thousands of enterprise sellers at AWS, GCP, and Azure are all effectively selling against OpenAI every single day. As Quinn put it: “When you walk into an account, you’re literally like — I will sell you anything but OpenAI. Because I can’t.”

Doom added some color: most of OpenAI’s go-to-market team are former AWS startup sellers and account managers. They know the motion, but they’re now the only sales force swimming against the current.

Until OpenAI gets its stateful platform (their equivalent of Claude Code and Cowork) running on other hyperscalers, the enterprise distribution problem is real.

Meta Launches Muse Spark

Meta Superintelligence Labs announced Muse Spark — the first frontier model built on their completely rebuilt AI stack. Alexander Wang shared on X: “Nine months ago we rebuilt our AI stack from scratch. New infrastructure, new architecture, new data pipelines. Muse Spark is the result.”

Quinn gave it a spin for image generation and wasn’t blown away, but the signal is clear: Meta is spending billions, the stack is new, and the competition isn’t slowing down.

Are Tokens the New Cloud?

This was the big question of the week, prompted by a16z’s excellent “Charts of the Week” on the SaaSpocalypse. The data from the Jefferies CIO survey is striking:

In August 2024, fewer than 10% of companies were spending 5% or more of their tech budget on AI. By January 2026, that figure jumped to over 60%. And the breakdown across spending tiers — 5-10%, 10-15%, and 15%+ — shows that this isn’t just dipping a toe. Companies with billion-dollar tech budgets are now committing meaningful percentages to AI tokens and infrastructure.

As Quinn put it: “Doom, are tokens the new cloud?”

Doom’s answer: “I mean, this says it is.”

The SaaSpocalypse: Nuance Over Narrative

We’re not on the SaaS apocalypse bandwagon — but we’re not ignoring the data either.

The a16z charts break down year-over-year spend changes among the largest AI-adopting companies. HubSpot is the clear winner, with the highest median and average growth in the panel. Figma, Box, Cloudflare, Datadog, and Semrush also posted strong 25%+ growth among the biggest spenders.

On the other side: Dropbox is flat-to-negative. Atlassian is down 21%. Asana is down 45%.

There was also a telling signal: 71% of CIOs expect to cut systems integrators to fund AI or cybersecurity. But zero respondents planned to cut cybersecurity budgets. Zero.

Doom made the observation that the OG incumbents — your Salesforces, your tightly-integrated stacks — are better positioned than the mid-tier point solutions. His team just added Salesforce seats. But he’s hungry for the Cowork MCP integration so non-engineers on his team can leverage Claude directly against their CRM data.

Quinn’s take: SaaS won’t die, but procurement is going to come in hard. Expect seat count pressure, base-plan downgrades, and revenue headwinds for any platform that isn’t going all-in on AI.

AWS Trainium Backlog

A quick signal worth noting: AWS CEO Matt Garman mentioned that all currently active Trainium infrastructure is already spoken for — and future capacity is earmarked as well. For context, Trainium is Amazon’s custom silicon designed for deep learning workloads with strong price-performance. It’s now also handling inference, which overlaps with their Inferentia chip. The backlog is likely multi-billion dollars, driven heavily by Anthropic’s demand.

As Doom put it: “Long story short — buy your call options for Amazon.”

Builder Corner: What We’ve Been Building

Quinn: Set up CI/CD pipelines pushing from GitHub to a hyperscaler running database, container, and gen AI environments — no manual laptop auth needed. Also configured Cowork scheduled tasks to process podcast transcripts and generate research prep docs using Opus 4.6. Created and presented an Enterprise Security Stack presentation (now on YouTube). Next up: a new category presentation every week.

Doom: Been using Dune Analytics with MCP integration to query on-chain data dashboards built by data scientists, then asking Claude to go deeper on unfamiliar metrics. The analysis comes back confident — sometimes too confident — so everything gets annotated “Hey, this is from Claude” before it goes to the team.

Quinn’s tip: Use Apple’s voice memos for quick brain dumps. The built-in Whisper transcription is rough, but paste it into an LLM and it cleans up instantly. Great way to capture authentic voice as a writing starting point.

Fringe Lines drops weekly. Subscribe so you don’t miss the next one. And if something hit different this week, tell us in the comments — we read every one. — Quinn & Doom

Subscribe now

4/ Recommended Tables & Nano Banana Pro Image Prompts

Recommended Tables

Table A — SaaS Spend Winners & Losers Among Large AI Adopters

Table B — AI Share of Enterprise Tech Budgets Over Time

Source: a16z / Jefferies CIO Survey panel of large AI spenders

Links & Resources

• a16z Charts of the Week: SaaSpocalypse Interrupted —

• Anthropic Claude Mythos Security concerns & Project Glasswing

• Anthropic Managed Agents

  ---

Subscribe now

It used to be that ideas were cheap and execution was hard. That equation is flipping fast — and what it means for enterprise sales, SaaS, and your career is worth paying attention to.

This week, we’re covering what we’re seeing on the ground: non-technical founders building entire back-office systems in a weekend, Goldman Sachs projections reshaping the SaaS TAM, why your customers want AI from vendors they already trust, and an open-source agent called Open Claw that might keep you up at night.

Subscribe now

A Tree Guy Built His Own ERP in 40 Hours

A buddy who runs a tree service business — no coding experience whatsoever — sat down for a marathon weekend and built a full back-office ERP system. He started on Claude Desktop, hit the context window limits, moved to the web interface with multiple tabs, then discovered Claude Code. He paired it with Whisper for voice-to-text input and spent the last 15 hours of his build just talking to Claude Code. The result? A working system backed by Supabase.

Meanwhile, we deployed a fully automated image generation stack on AWS using Amazon Kiro — cloud formation deploys it, sets up all resources, and it parses rows from an Excel file as individual prompts, sends them to Amazon Nova Canvas, and saves the outputs to S3. Swap in your preferred image or video model and you’ve got a production content pipeline.

The takeaway that keeps echoing: execution is getting really easy. The idea might be the hard part now.

Marc Andreessen made a related point recently — Claude Code is amazing, and it was built in two weeks. But flip that around: it only took two weeks to build, which means competitors are right behind. Distribution is the moat, not the build.

AI Hasn’t Made Us Better Sellers Yet — But It’s Changing How We Work

Here’s an honest admission from the trenches of enterprise B2B sales: generative AI isn’t dramatically improving the selling motion yet. Not for account managers going deep on a handful of named accounts. It’s not the SDR-scaling, outbound-automating revolution that the hype cycle promised.

Where it is helping? The internal bureaucracy. The paperwork. The analysis you never had time to do.

One practical example: spinning up an internal Slack channel called “AI for BI” and just dumping Excel files into Claude. Prompting it with “pull out five interesting anecdotes from this data” and getting instant signals — stack-ranked insights with actionable suggestions. Would you have done that analysis manually? Probably not. Would it have required a formal request to your data team? Definitely.

Another example: a finance professional watched someone use Claude to generate a cash flow statement from a balance sheet and income statement. Work that would have cost thousands of dollars in consulting fees — done in minutes. And honestly? The consultant would have hated doing it.

The pattern emerging: AI isn’t replacing the selling. It’s eliminating the friction around it.

Your Customers Want AI From Their Existing Vendors

This might be the most important signal for GTM professionals right now. From the OnlyCFO 2026 budget analysis:

“I don’t want to purchase from another vendor. I have too many apps already. I want to purchase AI stuff from my existing vendors. I bought them because I trust them.”

If you’re an enterprise seller, this is your green light: lead with your existing customer relationships. Talk to your customers about your company’s AI offerings. Don’t put your customer success managers up against hungry, hyper-native AI startup reps in what feels like an entirely new sale — they will lose that fight.

Win where you have your moats. Start from your strengths.

And to the people predicting the death of SaaS — the reality is more nuanced. Nobody is ripping out a thousand-integration CRM like Salesforce to replace it with a vibe-coded internal tool. Most of the cost of software is in the maintenance, not the building. Compliance, updates, security — that’s where the money goes, and that’s not going away.

The Goldman Sachs TAM Projection That Should Get Your Attention

Goldman Sachs estimates the total addressable market for SaaS at roughly $30 billion today. By 2030, they project that shrinks by about a third to $20 billion. Sounds scary — until you see the other side of the ledger.

The agent TAM goes from essentially zero in 2026 to an estimated $30 billion by 2030. Add them together and the total market grows to $50 billion. There’s more value out there, not less. The nomenclature is shifting from seat-based licensing to outcome-based models, but the opportunity is expanding.

SaaS is under pressure — but perpetual licensing was supposed to be dead for years, and they still use on-prem Oracle databases. Don’t write the obituary yet.

One spicy take we heard: enterprise software won’t die, but Excel and PowerPoint will. Not in function — you’ll still work with tabular data and presentations — but in form. You’re not opening a spreadsheet and manually formatting columns anymore. You’re dropping the file into Claude and asking questions.

Open Claw: The Personal AI Agent That’s Both Exciting and Terrifying

Open Claw (formerly known by several other names before trademark issues intervened) is an open-source personal assistant that orchestrates AI agents. Here’s what makes it different from the enterprise agent frameworks like CrewAI, LangGraph, or Amazon Agent Strands:

• It runs locally — even on an old Windows laptop or a Raspberry Pi

• It connects to your model of choice via API

• It communicates through a Discord server — you just message it and it works in the background

• It gives you updates when tasks are complete

One person in our network has it running on a 10-year-old laptop, connected to Claude’s API, taking instructions through Discord. It does what you’d normally do in Claude Code — thinking, executing, iterating — but autonomously in the background.

The promise is enormous. The risk is real: people have been hacked after sharing API keys on public instances. Prompt injection remains a serious vulnerability if it’s exposed to the internet.

But here’s the thought experiment that keeps us up: if one person can run one Open Claw agent, why not two? Or five? You treat each one like a separate employee with its own credentials, workspace, and isolation. At that point, what does headcount even mean?

The scariest and most exciting question: if your Open Claw and my Open Claw can handle discovery calls, share architectural diagrams, break down infrastructure portfolios, and match up partner solutions — when do humans step back in? Probably at the $1 million signature line. But everything before that?

The Macro Backdrop: Revenue Per Employee and the Coming Squeeze

We’d be ignoring reality if we didn’t mention the broader context. Unemployment is at its highest since the Great Financial Crisis. Major tech companies — Amazon, Microsoft, Google — are all laying off. You can argue that’s COVID correction rather than AI displacement, but the structural shift underneath is real.

AI is deflationary for software. It’s easier and cheaper to build. The top decile AI companies are running average revenue per employee around $750,000. If your business model was built on the old P&L of high-margin seat-based software, you’re now competing against teams that do more with fewer people — and they’re going to push prices down to capture market share.

The optimistic view: once companies right-size their ratios and restructure their business models, there should be room to climb back up the value chain. More people should eventually equal more output and differentiation. But the transition period? That’s where we are now, and it’s going to be bumpy.

What We’re Watching

• The AI GTM 100: A curated list of the top AI-powered GTM tools. Clay remains the only unicorn (valued over $1B). We’re tracking names like Air Ops (generative engine optimization), Attio ($116M raised, CRM play), and Unify (outbound automation). The question: are we entering the age of “lifestyle SaaS” where it’s so easy to build that you don’t need to be a $10M+ company?

• Content generation pipelines: We deployed a system using n8n that scrapes a fashion catalog, pulls all product images, and generates 8-second model videos using Google’s Veo 3.1 — all within an hour. Gamma Studio just upgraded, and the gap between “I had an idea for marketing content” and “here’s the finished asset” is collapsing.

• Product-led growth ceilings: Anthropic’s growth is remarkable — people are just requesting seats organically, no AE required. But remember Andy Jassy’s reflection: “I wish I had hired an enterprise sales team sooner.” PLG has a ceiling. Even rocket ships eventually need account executives.

The Bottom Line

If you’re in GTM, the message is clear: build, experiment, and protect your existing customer relationships. The tools are absurdly powerful. The competitive landscape is shifting fast. And the value — both the TAM projections and the day-to-day efficiency gains — is real, even if the perfect AI-native sales motion hasn’t been figured out yet.

Nobody’s cracked the code on AI for enterprise B2B sales. That’s not a problem — that’s an opportunity.

Until next week — keep building, keep selling, and keep asking Claude, ChatGPT, and Gemini the questions you’d never have time to research yourself.

Subscribe now