• Nvidia just launched Open Secure AI Alliance that argues for unregulated use of open weight frontier AI models, including the Chinese ones accused of distilling Anthropic’s models. Interestingly, Nvidia claims that releasing AGI-grade open weight models in the wild fosters AI safety.

• Relatedly, Satya Nadella wrote a series of Twitter essays on why enterprises should avoid working with Frontier AI models to protect their “data IP”, and train its own AI models by finetuning open weight models.

This growing rift among the U.S. AI ecosystem players is highly concerning, especially from the geopolitics angle. It creates dissent among American business leaders, which plays right into China’s hands.

But from a business perspective, it’s fascinating to inspect the motives of every player in this feud between “Nvidia-Microsoft-Application Software” and “Anthropic-Amazon-AMD-Google” camps. In some sense, this war between “frontier labs” versus “Nvidia-sphere” was inevitable, since they need to commoditize each other to survive (as we will see why).

And like it or not, how this feud resolves will impact the future shape of global AI ecosystem, as well as the AI capex spend, and ultimately all capital markets! So while I hate writing about trendy AI gossip, I’ll make an exception this time. I’ll also preface this post with a trigger warning, since “open weight AI” has become more polarizing than politics or Lionel Messi for some.

Read more

If you are an enterprise AI founder, when should you sell your company?

Recently, Salesforce announced it’s buying Fin (formerly known as Intercom) for $3.6 billion, implying a valuation of roughly < 9x “ARR” based on the $400m/yr run rate reported by Fin back in March.

I put “ARR” in quotation marks since $100m out of Intercom’s $400m run-rate comes from AI customer support, which uses outcome based pricing as opposed to traditional SaaS pricing. Anyone who’s sold CX (customer experience) software before should know that outcome based revenue is highly unpredictable - and therefore lower quality - due to seasonality and other whims of the market.

In any case, this deal is interesting particularly because of just how “low” the valuation feels at < 9x “ARR”, considering the stratospheric valuations of Intercom’s comps in Sierra or Decagon, which recently raised at $15bn and $4.5bn respectively at lower revenue run rates, implying > 50x “ARR” multiples.

Of course, 9x “ARR” is crazy high for a company like Intercom, especially given the downtrodden multiples public SaaS, so it’s obviously a great outcome for Intercom’s founders. It just feels very low compared to its private market comps.

So why did this deal happen in the first place, and what does this really imply for the rest of the enterprise AI market, not just voice AI?

Intercom’s acquisition is important for price discovery in private markets, because until now investors just marked up everything up as if startups are Palm Beach real state. The bet was that you could dress up a consulting business as an application layer play, and exit at software multiples. But what the heck is going on here with Intercom and Salesforce specifically?

Commentary on the Intercom / Fin Deal

First, the deal itself, and what each party gets out of it, strategically.

Overall, there’s a high likelihood that Intercom’s founders were trying to get acquired as early as Q1 2026, which goes against their big rebrand to Fin in May and going “all-in” on voice AI.

That’s because Intercom did a debt financing of $250m back in March, which is only logical if the founding team didn’t want any dilution ahead of a potential acquisition.

Upgrade to access premium content. See testimonials.

Join 60K monthly readers to stay ahead of the curve. Read by strategy, VC, product, revenue, and tech leaders at OpenAI, Anthropic, Gartner, Deloitte, AWS, Google, KPMG, Harvey, Sierra, Slalom, etc.

Subscribe now

So it’s possible that Salesforce wanted to acquire Intercom / Fin for some time, and it makes perfect sense, because:

Read more

Jokes aside, AI cost management — and more broadly, financial planning around token spend — are fast becoming massive problem areas for enterprises. At least, Zuck feels that pain, and Meta is reportedly building internal tools to track AI token spends firm-wide to better measure ROI on AI spend. Meanwhile, Uber is introducing per-employee budget of $1,500 a month, indicating that AI spend policies will have hierarchical limits (per business unit, per employee).

What’s funny is, these moves from Uber and Meta remind me, a former trader, of how risk management is done in Wall St banks and hedge funds. In finance, the “ingredient” to make money is essentially just… money. So it’s important to watch it like a hawk to prevent traders from blowing up your firm. Similarly, as AI takes increasing share of firm’s final output, tech industry will soon learn to do the same for AI tokens.

For example, if you are a stock trader at a hedge fund, you have $X million that you can lose each week, month, quarter, etc. If you lose more than that, they keep you on a tighter leash. If you blow through your risk limits too often without showing any results (making money, which is foreign to many tech workers), you are let go. If you try to mess with your risk limits by committing fraud, you are fired on the spot.

And in some sense, managing tech bros using AI agents for coding has deep parallels with managing traders, and the whole field sorely needs some concept of “risk limits” to constrain their behavior regarding AI usage.

Here’s how to think about this. Essentially, every developer using AI is suddenly a capital allocator (or a trader), since tokens are capital. More broadly, as “AI tokens” take up a greater fraction of a firm’s end output (versus labor), then every worker becomes a capital allocator.

So, if you are a business, you are effectively hiring developers as discretionary traders to make money with the capital stock that you are providing. Without risk limits or policies or compliance, you get an unconstrained shoveling of money into OpenAI or Anthropic’s coffers, straight from your wallet.

Unfortunately, companies are still mistakenly seeing AI agents as yet another IT spend. It’s NOT. It should be viewed as a capital allocation decision, especially vis a vis outsourcing. Arguably, you can’t put a clean, fixed budget on a spend that can increase geometrically and unpredictably. Instead, things should be aligned with work.

If the Wall St analogy holds, every enterprise’s FinOps need to be upgraded with things like:

• fat finger protection: prevent a rogue employee from running 50 hour background jobs with the latest fancy token sucking feature in Codex

• P&L measurement: measuring how much value was delivered for an agentic workflow that costs $X to run.

• compliance and fraud protection: as tokens get tied directly to real dollars, you need guardrails against abuse — shadow AI, shadow API keys, employees burning company tokens on personal side projects, or data quietly leaking out through prompts.

None of this is exotic. It’s the same middle-office plumbing every trading floor already runs, and the moment enterprises decide they need it, they will adopt it. We’ve seen this movie before. It was called cloud Fin Ops.

Note, if you are an enterprise that’s spending $10M+ on tokens and need cost optimization help, DM me as I have developed a framework to slash costs dramatically. Or stay subscribed for more info in the coming weeks.

Subscribe now

Anyways, between 2018 and 2023, a hot topic in enterprise cloud was “cloud cost optimization / cloud FinOps”, and a slew of companies were founded to help enterprises lower their AWS / GCP / Azure bills in exchange for some cut. Almost none of them exited for more than $1Bn, but still many decent businesses were founded helping companies save money on cloud.

For example, say you are a Fortune 100 bank, and you are spending $80M a year on S3 (AWS object storage) storing petabytes of data that’s growing 30% a year. To mitigate this problem, you hire a company to scan every S3 bucket, and detect / fix any wasteful configurations, such as storing archive data in expensive storage tiers. The cost savings were immediate and measurable, which is why “cloud cost optimization” startups did well for a while, until their feature sets got absorbed (read: copied) into AWS’s roadmap.

Fast forward to 2026, I think there’s a pretty wide-open market for providing products and services around “AI cost optimization” as well as “AI FinOps”, given the recent hoopla about exploding AI costs. Which is probably a premature worry, honestly, given that AI tokens represent less than 2% of total expense even at the most “AI-native” companies.

But it’s inevitable that CFOs start demanding tighter governance and monitoring around AI-token costs, and feed that data back into financial planning, especially as AI adoption increases over time.

This necessitates better tools to monitor and optimize AI token bills, as well as generate real time ROI calculations to assess the value of every token used inside the company.

For example, right now, no company is currently able to answer basic questions such as:

• which business unit is delivering the most “value”, adjusted for AI token spend?

• which models account for the most $ nominal token spend? Are we inadvertently using Opus 4.8 to do something basic like writing unit tests?

• how do we prevent runaway spend *before* the bill lands, instead of explaining it after the fact?

So what does the AI FinOps stack actually end up looking like? My bet is it splits the same way Wall Street and cloud both did.

On one side, the “risk and P&L” layer — the unglamorous, CFO-facing system of record that maps every token to a business unit, an owner, and a workflow, then ties it back to the value it produced, so finance can decide who gets more rope and who gets less.

On the other side, the execution layer — smart compaction, prompt compression, model routing and hot-swapping, ideally running right at the inference layer. Think of it as smart order routing in stock exchanges but for AI: genuinely valuable, but the kind of thing that gets commoditized or absorbed into the platform, the way AWS quietly ate the cloud-optimization startups.

But my advice to enterprises is, stop worrying about trying to do everything at once. Instead focus your efforts on enforcing basic hard limits per employee, taking an inventory of all API keys issued for every business unit, and visualizing how much every business unit is spending in aggregate on a trailing 30 day basis.

Everyone is interpreting Meta, Amazon, and Microsoft’s pushback on AI token spending as an evidence of AI not delivering value. I think that’s the wrong framing. AI and capital are mostly amplifiers. If you give money to a bad trader, he or she will lose money, and that’s the same with AI tokens.

The winners won’t be the companies that spent the least; they’ll be the ones that treated tokens like capital, built themselves a real risk desk for AI.

About Me

I write the “Enterprise AI Trends” newsletter (read by over 60K readers worldwide per month). Previously, I was a Generative AI architect at AWS, an early PM at Alexa, and the Head of Volatility Index Trading at Morgan Stanley. I studied CS and Math at Stanford (BS, MS).

After all, if you have literal AGI in your pocket in the form of Claude (a statement that many offend many of you), how can you not know what needs to be done, and how to do it? Why are 52 year old CIOs paying ex-junior-consultant-turned-FDEs to tell them how to run their businesses with AI?

But it turns out, almost halfway into 2026, most companies are still struggling to form a coherent AI strategy, let alone operationalize a meaningful AI initiative.

Why? Because the value of agentic AI is highly dependent on the talent and domain expertise of the users. Just imagine deploying Claude at your local DMV - it’s easy to see why AI tools alone won’t magically make companies run more efficiently. Give AGI to a random org with no incentive to change, and you get a very expensive chatbot.

Which means, there’s a huge “skill issue” when it comes to AI adoption, and it’s not just normal enterprises that struggle with translating AI tokens into real value, but tech companies as well, which are incinerating money as we speak.

• Consider how Amazon, supposedly, blew $500m in one month on Claude without seeing much ROI and had to stop token leaderboards.

• Uber blew its 2026 AI budget in 4 months and blamed Anthropic for it, without moving any of their business KPIs. Why? Because agentic AI usage skews to automation, not delivering more de-novo value to customers.

• Most startups are also full of it when it comes to using AI. For example, I know a non-technical founder who got “AI psychosis’ed” into thinking he’s a developer, and fired his CTO to run engineering, only to blow out his production database with his vibe coded slop. It’s a real mental disease.

So if AI adoption is a game of skill like poker, then you would expect some kind of Bell Curve in AI adoption progress. Thus, there are many companies that are crushing it with AI also. But they tend to be culturally excellent, talent dense teams with real incentives to increase productivity. And those folks aren’t the ones bragging about tokenmaxxing on LinkedIn. They are too busy making money, and would prefer to keep their edge on the DL.

So anyways, the playbooks are all there, there’s a ton of free info out there on doing anything, and it’s never been easier to write software. But companies still struggle to adopt AI. How? There’s roughly three failure modes:

• inaction: many companies still have no concrete plans to modernize basic things like AI assisted customer support or healthcare revenue cycle management

• overzealous chase of vanity metrics like # tokens used: there’s an unhealthy obsession with vanity-metric masturbation affects startups and big tech alike. But even the nerds at Meta are realizing how dumb it is to brag about lighting money on fire, and pulling back on token leaderboards.

• poor build / buy strategy: companies are building stuff that they shouldn’t be building to chase promotions (“resume driven development”) and not building things that lead to business value cuz it’d stir the pot too much.

Case in point: Kirkland & Ellis LLP (the world’s largest law firm by revenue) announced it will spend $500M developing its own AI platform — presumably a custom application (like Harvey or Legora), plus infrastructure to train custom models.

Read more

Doesn’t matter what domain you are in, every software company seems to have converged on a rather uninspired north star of becoming a command center for creating and running background agents for your company.

That explains why Cursor (coding), Linear (issue tracking), OpenEvidence (clinical questions), Perplexity Computer, OpenAI Workspace Agents, Google Enterprise Agent, Tasklet, and hundreds of other products look like… Claude Enterprise.

They all look something like this.

Notably, everyone has a giant text box where you are supposed to write your prompt, and upload documents and context. These products all feel very “minimalistic”. Why? Because adding more GUI elements will distract you from tapping into platform value (AI agents). We want to draw a straight line from the user to the agent.

But why stop there? Let users turn their productive conversations into repeatable automation templates, which can run in the background or on-demand. It’s like having an elastic personal fleet of minions. This turns work into a game of automating your human coworkers, until the org chart collapses, leaving just you, your buddies, and a thousand agents. Brilliant! If only you figured out that concept…

So why is this convergence of apps happening? It’s not because software companies are unimaginative copycats (ok, maybe some of them are), and couldn’t conjure up more creative UX.

It’s because AI has rendered software into basically power outlets / taps / faucets / gas stations for consuming agentic intelligence. If LLMs / AI is basically “intelligence on a tap / switch”, then the AI business is basically about buying & selling electricity, but in the digital domain.

And that explains the convergence of product features and UX that has happened this year. Ask yourself: how much innovation or variety do you see in the design of switches or faucets? Do you care about implementation details like how wires are laid out under it? You may be willing to pay extra for a “smart outlet”, but that’s not everyone.

The takeaway: product layer is getting thinner, and ultimately users will just want a simple interface to accessing power, unless there’s a real modality mismatch between the AI and the user experience (you can think of Robotics or Voice AI as a valuable “GUI” in the LLM era).

Read more

Read more

Subscribe now

I know no one’s interested in law, let alone legal AI, but it’s growing really fast and we need to talk about it.

It seems like every other day, Anthropic or OpenAI ships a new vertical “skill” or plugin—and the basket of stocks in that domain gets a jump scare.

A few weeks back, Anthropic launched a skills marketplace and a legal contract review skill, and Thomson Reuters and RELX sold off hard. And it’s not just legal: when Anthropic shipped code review capabilities, CrowdStrike and other cyber names got clipped too. You can almost sense that trading algos are trading off of Anthropic and OpenAI’s blog.

Overreaction? Maybe. For one, these skills won’t replace a trained lawyer on their own. Liability, client trust, and regulatory nuance aren’t things you can just prompt as a layman. And “lawyer” isn’t even a single job—it’s a bundle of jobs-to-be-done wrapped in CYA (liability management).

But the market’s instinct isn’t crazy. In my view, AI can hit services harder than software—especially anything priced in billable hours. Not because “AI replaces lawyers” tomorrow. Because the meme is right: it’s not AI that replaces you, it’s another person using AI.

Read more

Which means, presumably, many are probably thinking about being “contrarian” and start buying software companies again.

But that’s not as contrarian or smart as you might think, either: trying to catch falling knives is a pretty common mistake. So unless you are good at trading, you may get whipsawed in the coming months of volatility.

Rather, the path forward is more dispersion: the market will keep punishing losers, while those that survive will be labeled as “winners”, but only in retrospect. The alpha is knowing which is which, in real time.

So in this post, I’ll share my thoughts on what public software stocks might outperform and have the best chance at surviving the transition to AI, after re-listening to 50+ calls and triangulating to the latest trends. Also, I’ll also provide a 10-point checklist to rank long / shorts within SaaS space.

But first, a word of real caution to the dip buyers: buying beat-down stocks can be very risky when there’s an industry-wide malaise. I’ll give you an example.

Back in 2008, many thought financial stocks bottomed when Bear Stearns collapsed and got bought out by JP Morgan on March 16th, 2008. But another, bigger crash ensued when Lehman Brothers collapsed on September 15th, 2008. The real bottom in S&P was the following year. Over 400 banks failed from 2008 to 2011.

Share

Now, I’m not saying SaaS industry is the equivalent of mortgage or banking industry in 2008. Most software companies won’t go under, but rather turn into low growth zombies, especially since they aren’t levered. Many will turn into the new toilet paper or newspaper company-equivalents. What I’m saying is:

Read more

Share

I’ve been thinking about how to explain Claude Code / CoWork better to people. The “Excel” analogy is the first one that actually sticks.

For forty years, Excel was the undisputed king of general-purpose knowledge work. It separated the savvy from the rest in every white-collar function on the planet. It became the lingua franca of finance and accounting, and then quietly colonized every domain. Entire businesses still run on spreadsheets attached to emails.

But that reign is ending, and people aren’t paying enough attention.

Agentic AI tools—Claude Code, CoWork, OpenAI Frontier—are on track to replace Excel as the baseline skill of knowledge work.

Not “someday”, but this year.

And just like people ran entire businesses on spreadsheets, future businesses will run on mostly on whichever wins the agentic OS war. Claude Code already runs my business in every functional domain.

In that world, vibe coding is what pivot tables were: a higher-order fluency that lets you get more leverage inside the same tool everyone has. What was “coding” before becomes model training and finetuning, especially since the “model” IS the “product”.

But the market is still mis-pricing the coming ubiquity and impact of Claude Code (or OpenAI Frontier, used interchangeably here). Anthropic dropped an interesting stat yesterday: Claude Code is already at roughly a $2.5B run rate, growing 100%+ in January alone.

That implies a product that will transcend any particular domain boundary, aka it won’t be contained to just coding. Also implied is a huge head room remaining. Excel is a ~$40B business. Office is roughly ~$120B. That gap alone suggests Claude Code can still grow 50x. It’s still very early.

What matters isn’t the current number; it’s the slope.

Claude Code is compressing what took Excel decades into quarters through viral PLG. The DevRel playbooks on social media are the well-oiled machines of spreading Claude Code through productivity porn.

And when something becomes a standard, the ecosystem reorganizes around it. Excel didn’t just win because it was a spreadsheet. It won because it became the default interface for “doing work with data.” Once it owned that position, everyone plugged into the Excel ecosystem, not the other way around. Excel became an assumption.

In other words, Excel is / was the substrate of work. Similarly, the industry will organize around the new “Suns” (Anthropic / OpenAI), with skills, plugins, and integrations orbiting the core tool. People are already frantically contributing Claude Skills into shared repos.

That has implications most enterprise AI builders are not ready for. For one, it becomes clear what are the “tarpit” ideas. Nobody built “Excel for accounting” or “Excel for inventory management” or “Excel for Financial Modeling”. People learned Excel and that was that. Excel captured almost all value in that space.

The software that survived wasn’t “more spreadsheet.” It was software that had real structural differentiation over Excel—especially around UX, coordination, workflow, and proprietary data.

But it will be harder to differentiate from Claude Code than it ever was to differentiate from Excel, because Claude Code is more general-purpose, more composable, more scalable, and improving every day, not at some release cycle cadence. UX is less meaningful when agent delivers outcomes. Thus, recent UX innovations are happening mostly for alleviating the pain of managing agent swarms, etc.

Speaking of vibe coding, many tax and finance professionals I personally know are vibe-coding lightweight apps already, with zero coding background. Some are even launching their own AI powered CPA services, etc. The constraint is domain knowledge now, not web development skills.

Now, before some of you cry “but, but it’s not production quality“ or “non-developers don’t know what they are doing, they will hire a developer to fix their vibe coded mess”.. here’s what many don’t get:

• Businesses run on a ton of “brittle spreadsheets” already. People have been “vibe spreadsheeting” for a long time. Vibe coded apps are an extension of that, except Claude Code is already super-human in its ability to whip out simple dashboards and CRUD apps.

• Most enterprises have 1-2 gung-ho people (often young people eyeing promotions) who will go to great lengths to build automations. When I started as a trading analyst back in 2007 (great timing, just before the financial crisis), I spent almost a year building and operating spreadsheets to help senior traders. That grunt work was what got me promoted to associate. Similarly, people that use AI effectively at work will be the new rockstars.

So when is this shift going to happen, you ask. It’s already happening.

Claude is eating Excel jobs, and finance industry is the canary in the coalmine. Wall Street is adopting Claude both top down and bottom up. Goldman Sachs reportedly embedded Anthropic engineers for six months to automate large back-office functions—accounting, compliance, KYC—at real scale (confirmed via their earnings call). Even slow-moving banks are reorganizing around Claude as the operating system. By 2027, every verifiable task will follow finance and accounting trajectory.

But the analogy between Excel and Claude Code ends there.

Claude Code is far more disruptive than Excel, and creates a K-shaped divide along the fault line of AI literacy.

• Agents give you outcomes, not just automation. Excel automated calculations. Agents complete entire tasks end-to-end—draft the memo, review the contract, file the report.

• Agents scale through self-multiplication. Spin up N agents doing N tasks in parallel. You’re not one person working faster. You’re one person with a team.

• Agents self-improve without you doing anything. Every time Anthropic or OpenAI ships a new model, your agents get better overnight. Your spreadsheet never got smarter on its own.

Will everyone learn how to use AI? No, just like the majority of the population never figured out Excel. But the impact of not using AI will be devastating for many, and accelerate K-shaped economy. Not knowing Excel meant you were less competitive—slower, more manual, but survivable. Not using AI agents means you’re almost bionically outmatched, since you are competing with cybernetically augmented people.

Subscribe now

Implications for Enterprise AI

Right now, there’s a massive whitespace to fill in enterprise AI. Much of it will be filled by Claude Code or OpenAI Frontier—if these are truly the “Excels” of this era.

The real kicker versus Excel: these platforms are self-improving. They keep absorbing use cases like a black hole. Six more months of 100% MoM growth and enterprise AI startups are selling into a fundamentally different landscape.

The faster Claude Code grows, the further the TAMs will shrink for startups.

Nobody launched “Excel for Accountants.” Accountants just learned Excel.

The software that survived solved problems Excel structurally could not easily—Salesforce for CRM, SAP for ERP, Bloomberg for real-time market data. Regulatory moats, proprietary data feeds, genuine domain complexity—those still matter. But the pluggable architectures of skills and plugins keep narrowing the gap, and computer use agents will close whatever’s left.

Just like no one wants to use Lotus 123 or Apple’s Numbers for spreadsheets, same will happen to enterprise AI platforms. It’s hard to imagine scenarios where an enterprise adopts a multitude of agentic platforms, especially when the trend is to consolidate subscriptions and IT spend.

This means startups will have to think harder about differentiation and execute faster than Anthropic’s or OpenAI’s teams, lest they get priced at service company multiples. Deep vertical data, regulatory moats, genuine domain complexity (drug discovery, surgical robotics), durable network effects—those are real moats. Everything else is a feature Claude Code ships in a quarter.

But the real question is this: is there really any moats left when most value is in the capital?

Join 60K readers per month to stay ahead of the curve on AI trends.

Premium subscribers will get access to exclusive content. Fun fact, many execs at the big labs are subbed. See testimonials.

Subscribe now

About Me

I write the “Enterprise AI Trends” newsletter (read by over 60K readers worldwide per month). Previously, I was a Generative AI architect at AWS, an early PM at Alexa, and the Head of Volatility Index Trading at Morgan Stanley. I studied CS and Math at Stanford (BS, MS).

And indeed, chaos was what ensued during last year’s “DeepSeek Moment.” Even though DeepSeek’s $6M TCO training cost claims turned out to be wildly misrepresented, Nvidia still dropped 17% in a single day, and a six month debate between the “Jevons Paradox”-camp and the Michael Burrys of the world followed.

Of course, we all know how that debate ended, at least for now, with OAI and Anthropic revenue surging to end 2025 at $30bn run rate, and hyperscalers announcing a re-upped $600bn capex spending plan for 2026.

But if you are a serious investor or a CTO, you should still get paranoid about getting China wrong again, even though the AI landscape has become a lot more legible since the simpler times of 2025. Especially since Twitter and LinkedIn thought-fluencers are totally quiet about this, suggesting that people are too complacent, or busy vibing CRUD apps on Claude Code.

So in this post, I’ll share my thoughts on the 1) base case, 2) mild bear case, and 3) real bear case around next week’s DeepSeek V4 launch, and more broadly, how to think about Chinese “open source” models from a strategic lens. Depending on what actually gets released next week, and how, there could be positive or negative market reactions.

As a neutral commentator, the goal here is to perform a scenario analysis of what headlines hit the tape, and how to react with a calm, collected mind, as opposed to running around like a headless chicken.

Join 55K monthly readers to stay ahead of the curve. Upgrade to access premium content. See testimonials. Prices go up again soon.

Subscribe now

The Only Variable That Matters

When these models drop next week, the question everyone will be asking is: has China caught up? And if so, what breaks?

That’s not the right question to ask.

Read more

I’ve been in NY for the past week (a big mistake, it’s freezing) seeing my Substack subs ranging from startup founders, system integrators, and AI transform firms.

And I noticed one common pattern. Everyone’s assuming that enterprise AI adoption will remain messy and hard due to integration pains (”customer’s data is a mess”) and education hurdles (”customers are clueless about AI let alone know how to build AI agents”). Hence, FDEs and EM / PMs are needed to manage engagements and handhold customers.

But things are changing so fast that these assumptions will age terribly over a 12-18 month horizon, even sooner. If that happens, it’s obvious where the value accrues. In fact, I am actively advising young engineers from becoming FDEs, which is fast becoming a program management job w/o meaningful technical challenge (unless you like that, of course, YMMV).

What many are missing:

• First, every incumbent vendor has woken up to the importance of the “last mile” of AI, and is spinning up FDE and white glove GTM programs as we speak. They are embracing services, aka “schlep” and bundling FDEs to sell their platforms. This means startups have a shrinking window where their GTM is a differentiator. Most startups also know this, which is why they are hiring more sales people than developers, since their products don’t sell themselves like Claude Code.

• Second, more importantly, new technology is being invented where AI adoption doesn’t involve enormous integration pains. More on that.

So what happened?

This past week’s been very eventful, but I’ve picked out two discussion points in this edition of weekender.

First, OpenAI launched Frontier and formally announced that they are bundling FDEs with a new enterprise AI agent platform called Frontier. It’s quite unlike anything I’ve seen on the market, and if it works as advertised, I believe it will absolutely destroy AI startups’ offerings regardless of the vertical domain.

This is the crux of our post today. Related: OpenAI’s enterprise AI strategy for 2026.

Second, KPMG accidentally wrote the playbook for compressing consulting margins to zero. Supposedly, KPMG asked its own BPO to “pass on its AI savings”, stiffing them of a few hundred grand. Which is funny, because KPMG itself is a service firm that uses AI. All this is pointing to services being toast.

Two different stories, same punchline: the FDE model is eating the world until there’s nothing left to be eaten.

Let’s get into it.

I am making this post available to all premium tier subscribers. Upgrade to premium to unlock this post. See Testimonials. Join 55K readers in staying ahead of the curve in enterprise AI.

Subscribe now

Gift this subscription to your employees so that everyone gets smarter, not just you.

Give a gift subscription

Read more

Join 55K readers monthly to stay ahead of the curve on Enterprise AI Trends

Subscribe now

As collateral damage from the 2026 software meltdown, consulting and outsourcing stocks have been getting hammered too. That’s unfortunate (ouch)... because a lot of buy side investors came into 2026 “long professional services”—names like Accenture, Gartner, Genpact, IBM, which were already beaten down throughout 2025. (2026 YTD performance: ACN -10%, IT -36%, IBM -1%, G -17%).

The bullish thesis was straightforward. If 2026 were to be the year of “AI transformation”, then the natural winners would be the consulting firms, since most enterprises opt to hire outside help. Firms like Accenture, Genpact, or Gartner were supposed to print money doing AI transformation projects (which they are!). Plus, consulting firms can dogfood AI tools to improve their own profit margins per engagement (also true!).

But that trade hasn’t worked out so well so far. Gartner sold off 20% after posting earnings (Feb 2nd), and Accenture and IBM are also down YTD after a head fake in January.

So is consulting / sthe next shoe to drop? Can these “professional services” stocks turn around, and what are the short term and long term effects of AI on these businesses?

The hand wavy answer is “AI is killing consulting”, but that’s not exactly true, at least yet. It’s like how software hasn’t seen the seat compression yet, but revenue is a backward looking metric. Similarly, the consulting firm CEOs all affirm our bull case: 1) AI disruption fears are increasing the demand for consulting, and 2) consulting firms are using AI heavily to improve their own margins (see Deloitte getting caught for using AI).

But when we take a step back and listen very carefully to what the management is saying, there’s some worrying trends developing beyond 2026.

First of all, none of these firms gave great guidance (except IBM, which has a reaccelerating mainframe business). But more importantly, there are signs that AI - while being a short term tailwind - can flip into a headwind, and growth may decelerate into 2027. I’ll go over my rationale in this post.

In the following premium section I’ll talk about my view on professional services firm valuations as well as offer a vision for “future of services” in the Claude Code era. Prices for regular paid plans will go up at the end of February.

Subscribe now

Read more

Subscribe now

Is n8n / low-code AI automation dead?

That’s been the elephant in the room for many, so let’s deal with it.

Over the last 3 months, Claude Code and CoWork introduced a new way to build AI automations (by writing markdown files), and it quietly broke a core assumption the workflow industry has lived on for years: that non-technical users need a visual, drag-and-drop builder to create reliable automation.

So this isn’t just about n8n, but about the design philosophy beneath a huge chunk of enterprise software and “AI-native” startups. ServiceNow’s workflow stack is built around designing “flows” and visual orchestration. LangChain has LangGraph Studio as a visual builder. Sierra and Decagon (customer support AI startups) also have a visual builder. Relevance AI, Flowise, Gumloop, etc, the list is long.

And over time, everyone’s products started looking more or less the same. Visual flow based builders were the consensus bet when it comes to enterprise AI automation.

But then Claude Code happened, and that bet is starting to look sketchy.

Not because visual builders don’t work—they do—but because the jobs they were hired to do can be solved arguably better by a different paradigm. When you can describe an automation in plain English and get deterministic logic back, the drag-and-drop GUI starts getting really annoying and cumbersome. Not to mention, n8n wasn’t meant for building AI agents with 50+ steps.

So if the visual builder was the core bet behind all these enterprise AI platforms, where does that leave us? Low code automation isn’t dead, but its role in enterprises’ tech stack is evolving quickly, and that’s the crux of this post.

This shift will affect everyone.

• If you’re a PM or operator, you’re wondering whether to learn low code automation, or go straight to Claude Code.

• If you’re in the C-suite, you’re wondering whether your AI transformation efforts are already a waste of money or whether that Sierra renewal was a mistake. You are wondering how to calm down your board members who read about Claude Code on Twitter.

• If you’re an investor, you start second guessing whether that ServiceNow long position makes sense, and if the consumption based low-code automation category (and the AI startups built on that paradigm) still has a future.

But as usual, the reality has a lot of nuances, and the transition from low code automation to natural language automation will happen in shades.

Join the 55K readers monthly to stay ahead of the curve on enterprise AI trends

Subscribe now

Read more

As someone who is allergic to hype, I went in hoping to rip it apart. And in its current form, it’s nowhere near primetime. The project requires a VPS, API keys, and manual security hardening — it’s really a developer toy right now.

And frankly, much of the “innovation” is packaging and marketing. You could already use Claude Code from a mobile app. Plenty of people had wired LLM agents to Telegram via n8n before OpenClaw existed. If history rhymes, there’ll be a wave of Mac Minis hitting eBay the moment the novelty wears off -- especially since the project was hastily rebranded from “Moltbot” over obvious copyright concerns.

But despite all that, OpenClaw may be the most important open-source AI project since Claude Code.

Not because the software is a technical breakthrough, but because it’s the first project that validated real demand for a specific product category that nobody’s built properly yet: ambient AI assistants.

More concretely, every major AI tool today -- Claude Code, ChatGPT, Copilot, Cowork -- assumes you’re in the loop as a manager. You sit down, open the tool, give instructions, supervise the work, and when you close your laptop, the AI stops. Every bit of Claude experience assumes a human to be in charge, and for a good reason.

OpenClaw flips that assumption. It trusts that the AI can act autonomously on its own, even when you’re asleep, away from your desk, or on a plane. That’s what “ambient, always-on AI” actually means: an assistant that doesn’t wait for you to open an app. It’s running 24/7, watching for things that matter, and taking action on your behalf.

And this sets the goalpost for where “AI assistants” from the incumbents need to be. Tens of thousands of people fought through real pain to install this thing, not because the software is exciting, but because they wanted to experience a new form factor for AI. OpenClaw proved the demand for what happens next: an ambient, proactive AI assistant with persistent memory about who you are.

That demand validation is going to wake up the incumbents - and who knows, Anthropic may launch a more polished OpenClaw next week.

But as more companies pile onto this “Claude Code Wrapper paradigm”, some clear mega trends will start emerging.

For one, the current AI infra stack ranging from observability to guardrails to protocols are not ready for a new Internet where AI agents are their own autonomous entities. They were built to assist humans understand the behavior of simple AI agents with maybe 20 tool calls. There’s significant work to be done in improving the tooling for agents that run 24/7.

And if we are heading in a direction where every person gets their own Clawdbot, we’re talking about changing how we use the Internet from a foundational level. That’s a platform shift across the entire Internet stack has not been discounted yet.

The obvious first-order insight is to be bullish on agent sandboxes and edge computing. But as I’ll argue, there are complications with that thesis -- and that the second- and third-order infrastructure plays are more interesting.

So in this essay, I’ll cover:

• What OpenClaw actually proves about AI’s next form factor

• Why incumbents -- especially Apple -- are best positioned to build the real version

• Why edge computing isn’t the answer yet

• The “agent exhaust” problem and why it drives massive storage demand

• Three missing infrastructure layers: flight recorders, runtime guardrails, and the agent-native internet

• The investment trades that follow

Join 55K readers per month on staying ahead of the Enterprise AI Trends.

Subscribe now

Consider supporting the publication and independent AI journalism by upgrading.

What OpenClaw Actually Proved

Before going further, let me level set on what OpenClaw actually is. This isn’t a comprehensive guide, just enough to understand what it proved and why it matters.

The simplest way to think about it: OpenClaw gave Claude Code its own computer and told it to act like a personal assistant.

If you’ve used Claude Code or Anthropic’s newer Cowork mode, you know the experience. You sit at a terminal, give it a task, watch it work. It’s powerful for software development -- but fundamentally synchronous and project-scoped (e.g. ~./claude). It runs on your machine, in your terminal, in a specific project directory. And when you close your laptop? Claude also stops.

Want AI to monitor your email while you sleep? Claude Code can’t do that (unless you use your laptop as a server). Need research surfaced while you’re in a meeting? Not its job. It was invented as a tool for doing work primarily, assuming a human in the driver’s seat, and a great one at that. But it’s not an always-on assistant.

The other gap is memory. You could technically use Claude Code with persistent memory (its CLAUDE.md files carry context across sessions) and even connect to it from your phone. But it wasn’t designed for autonomously building and managing memory about YOU over time.

So OpenClaw’s core innovation was surprisingly straightforward: give Claude Code its own computer to run on 24/7 and configure it to act as a personal assistant rather than a coding tool.

Originally created by Peter Steinberger — the entrepreneur behind PSPDFKit, one of the most widely used PDF tools — the project runs on any server you control: a cheap VPS on DigitalOcean or AWS, or a physical device like a Mac Mini. The intelligence comes from whatever LLM you connect. Most users plug in Claude from Anthropic, because Claude is still the strongest model for agentic tasks and has better guardrails against doing stupid things unprompted.

Now, the messaging integration (Telegram, WhatsApp, Discord, Slack) gets the most attention, but it’s not really the novel part. People had been wiring LLMs to chat apps for a while. You could already text Claude Code from your phone.

What’s genuinely different are two things:

• Persistent personal memory. Unlike Claude Code, which scopes its memory to a project directory, OpenClaw builds context about you across all conversations. It learns your preferences, priorities, and communication patterns over time. Claude Code knows your codebase. OpenClaw knows your life.

• Proactivity and ambience. Claude Code waits for you to type a command (though it supports scheduled tasks). OpenClaw is encouraged to initiate conversations when it notices something relevant, such as a calendar conflict, an email that needs attention, a task deadline approaching, etc. It can reach out to you without being asked, one of the benefits of AI living inside its own server.

Combined with having its own computer, meaning it can interact with GUIs, book flights, all things that terminals aren’t great for.. this creates an experience that feels genuinely different from anything else on the market.

To recap, OpenClaw’s product philosophy specifically indexed on four dimensions:

1. Ambience -- it works while you’re away from your desk, on tasks that have nothing to do with coding

2. Persistent memory -- it knows your full context across all conversations, building a model of your preferences and priorities over time

3. Computer use -- it can operate a full desktop environment, not just generate text in a chat window

4. Proactivity -- it reaches out to you when conditions change, without being asked

Now, this is all “happy path”. Bad surprises can happen (and have happened), and the security gaps and learning curve will eventually prevent OpenClaw from winning this category. There’s a big gap between validating demand and winning a market.

OpenClaw’s Critical Limitations

Read more

Vibe coding accelerates just one part of software development cycle — feature development. This creates enormous bottlenecks both upstream and downstream of coding, such as security reviews, SRE, and even product management. To truly lift outputs, every part of software development needs to be elevated and re-imagined for the AI native era.

Subscribe now

Vibe coding is going mainstream. Ryan Dahl (creator of NodeJS), Andrej Karpathy, and a growing number of graybeards have embraced spec-driven, vibe-first development. Enterprises are finally ready to adopt at scale.

But as execution speed increases, the software “supply chain” - from turning an idea into a PR, to reviewing it, securing it, shipping it, and operating it - is starting to break.

Most companies are framing this as a skills question: “What do we need to learn to move faster?” I think that’s the wrong frame. It’s not primarily a skill issue. It’s a supply chain and incentives issue.

In any production system, the final output is determined by the slowest stage. You can double capacity at one step and still see almost no improvement in finished output if the constraint moved elsewhere. The result looks like progress (more work-in-progress everywhere) while the actual thing you care about, i.e. shipped, reliable functionality, barely moves. “Work in progress” explodes.

This leads to a revelation: even if OpenAI or Anthropic “solves” coding, many organizations still won’t unlock the value—because their constraint won’t be coding anymore. It’ll be everything around coding.

You can see the shape of the problem by treating software delivery like a knowledge-work supply chain:

1. product & design (deciding what to build and what “good” looks like)

2. development (turning specs into code)

3. review & testing (quality, correctness, maintainability)

4. deployment (release mechanics, change management, rollback plans)

5. operations (reliability, monitoring, incident response, cost control)

Over the past decade, DevOps and microservices gave enterprises some real benefits, but also turned delivery into a distributed pipeline with many handoffs and many “humans in the loop” who own specific gates. That worked when coding was the slow part. It breaks when coding becomes fast.

Upstream, product managers and designers become the constraint: if you can implement ten times more than you can sensibly prioritize and specify, the backlog of “things we could build” turns into noise.

Downstream, reviewers and security teams become the constraint: if you can generate 150 PRs a week, but your org’s approval capacity is built for 20, work piles up. People route around the gate (“just merge it”), or they wait and the productivity gains evaporate.

The chip supply chain analogy is useful here. When AI chip demand exploded, the constraint wasn’t “making more GPUs” in the abstract. It was specific stages and components that didn’t scale at the same rate. You can ramp fab capacity, but if packaging capacity or HBM supply lags, finished systems don’t ship proportionally. The whole pipeline is paced by the slowest component.

Vibe coding creates the same dynamic: your shipping velocity is only as fast as the least-scaled stage in your delivery pipeline. And that stage is often a human approval process. AI increases “work creation.” It doesn’t automatically increase “output”.

In this post, I’ll walk through what that implies, and where the opportunities are.

Join 50K readers per month for the best thought leadership on enterprise AI.

DM me if you:

• need advice about adopting vibe coding at your enterprise

• any business inquiries for Enterprise AI Trends

Subscribe now

Why Enterprises Will Struggle (And Startups Can’t Afford To)

Startups aren’t advantaged just because they’re smaller. They’re advantaged because they’re allowed to rapidly change the shape of the manufacturing process.

They don’t have the luxury of preserving legacy gates, legacy ownership boundaries, and legacy incentives.

Enterprises, by contrast, spent years building elaborate change management processes: review rituals, approval committees, compliance gates, separation of duties.

Ten years of Microservices and DevOps evangelism preached by Netflix and AWS did a number. It aggravated data and knowledge silos, for one. AI agents work best when the context is all aggregated into one place.

But having 500+ internal code bases with some dependency over network complicates things for AI agents, simply because they can’t easily gather all the requisite information. That’s one advantage of monorepos.

Now, elaborate release management processes and SOA (service oriented architecture) are fine in a slower world. But organizations also learn to live inside their bureaucracy. They build kitchens designed for many chefs, and then act surprised when cooking takes longer.

When vibe coding hits an enterprise without a corresponding redesign of the downstream system, the “sacred” parts become the constraint. Code review is a perfect example.

Code review is still fundamentally human-driven. It’s the same handful of senior engineers who could barely keep up with 20 PRs a week. Now they’re staring at 150. PRs pile up, queues grow, lead time explodes.

Developers either start merging without review (risk spikes), or they wait (and the productivity gain disappears). You end up with the worst of both worlds: more change and less control.

And code review is just the beginning.

Security is full of manual gates: threat modeling, exception approvals, dependency risk decisions, and compliance sign-offs that don’t compress just because code generation got cheaper.

Testing often still depends on human QA for edge cases, exploratory validation, and acceptance criteria that weren’t crisply specified. Deployment can still require approvals from multiple teams because “who gets paged at 2am” is a political question, not a technical one.

Then you get the organizational multipliers: siloed tools (security and dev working in different systems), incentives that reward blocking (“not my job”), and blame culture that treats bottlenecks as safety (“if it breaks, at least it wasn’t on me”). Cultural inertia becomes a real issue (“we’ve always done it this way”). All of this prevents vibe coding to even get bootstrapped.

So the real enterprise failure mode isn’t that they can’t adopt vibe coding. It’s that they adopt it, generate a flood of output, and then discover their organization can’t stomach any real change.

Food for thought: do they even want faster change?

At that point in the future, there are only two ways forward, assuming AI keeps progressing: redesign the org chart and processes (the gates, change incentives, collapse handoffs), or reduce the work-in-progress by reducing people and scope, i.e., the ugly path of “lightening the ship” through layoffs and consolidation.

Startups don’t have that option set. Bottlenecks are existential, not annoying. They can’t wait weeks for approvals. They can’t ship broken code because QA couldn’t keep up. They’re forced—by survival—to re-architect their delivery loop around speed, clarity, and tight feedback.

But do enterprises have the luxury of working at the enterprise cadence, especially going forward? That remains to be seen. My conversations with CTOs indicate that things are changing fast in 2026.

Which raises the real question: if “writing code” stops being the limiting factor, what becomes the new scarce resource—and what do you rebuild first so the whole system can actually go faster?

Paid tier readers will have access to my tactical market commentary focusing on the biggest shifts in AI (like this), as well as my paid post archive. Support this publication by joining paid tier. Join hundreds in our community.

Subscribe now

The TAM Explosion for DevSecOps Automation

One consequence of this bottleneck problem is that there will be an explosive need to automate everything upstream and downstream of vibe coding.

By necessity, every mature company adopting vibe coding needs to automate review, security scanning, testing, and ops to avoid the bottleneck problem. Again, your processes move only as fast as your least automated bottleneck.

This will lead to an explosion in the TAM for better tools and services around DevSecOps. The demand will grow in lock-step with the progress in agentic models. The better the foundational models become, the greater the demand for process automation.

Interestingly, the TAM for pipeline automation tools scales with code volume, not headcount.

That’s a fundamentally different growth trajectory than traditional dev tools SaaS, which scale with team size.

In anticipation of this demand, the labs and the leading startups (e.g. Cursor, Cognition) have recently ramped up its offerings in DevSecOps. For example:

• OpenAI keeps investing into security agents for Codex because they know security bottlenecks block enterprise adoption. You can’t sell an enterprise CTO on “generate code 10x faster” if the response is “great, now I have 10x more code to security review and I don’t have the team for it.”

• Cursor acquired Greptile and launched Bugbot because they’re vertically integrating the toolchain. Greptile helps with codebase understanding and search—critical when you’re reviewing AI-generated code that touches unfamiliar parts of the stack. Bugbot automates debugging. They’re building the full pipeline: generate code, understand it, debug it, ship it. They can’t sell “10x code generation” if that code can’t ship safely.

• On a longer horizon, Cursor seems to be going for vertical integration of all parts of software development supply chain.

• In response, Cognition has also recently released an incremental upgrade to its code review offering.

In any case, the market map for “automating non-programming tasks” is taking shape. However, the adoption of “vibe coding adjecent” tools has still been incredibly slow in enterprise, mainly because of inertia. That’s why Cursor and Cognition - which are leaning more B2B and enterprise recently - are actively working on helping customers in this space.

In any case, labs like Anthropic and OpenAI will capture most value by bundling basic automation into their coding tools. Claude Code already ships with security review and code review agents.

However, the TAM is so large that there’s room for nimble startups to provide best in class offerings in all areas of the software supply chain, and be successful. I share my thoughts on this later in this post.

Some Tactical Advice for Enterprises

For companies that want greater feature velocity, start investing in automation right now, both upstream and downstream to vibe coding.

For small startups, they have the advantage of being unburdened by reputation risk and having large, heavy org charts and processes.

For the leading AI labs - which still retain much of startups’ advantages - they can vibe code solutions at breakneck speeds also. That’s why Anthropic shipped Claude Cowork in 10 days, which was 100% vibe coded.

But what about traditional enterprise? They are currently caught in a limbo where they are staring at a big pile of outdated processes, cultural issues, and technologies that are keeping them from shipping faster. The blockers are almost all unrelated to the performance of agentic AI models.

So here’s what I’d advise enterprises. Don’t do sweeping org changes yet, but focus on these three priorities to get the ball rolling and score quick wins.

Priority 1: Improve code review agent performance by investing heavily in documentation

Needless to say, code reviews need to be streamlined, but with a catch. Humans in the loop are necessary for approving the important PRs. 100% automation of code reviews is not desirable or possible, because human developers need to retain an understanding of the code base.

But that doesn’t mean the reviewer needs to read every line of code written by AI. AI can make code reviews more pleasant by providing better UXs, summarizations, etc, so that not every line of code needs to be eyeballed. There’s been a ton of new products launched just to streamline code review - take advantage of them.

Now, for code review agents to be effective, your team has to put in the work to teach AI agents about your company/domain specific preferences. That’s where the real work is - training the AI.

Out of the box, AI agents won’t understand your company-specific terms, like what AbstractCQASequencerController means (I made that up). One effective approach is to leverage LLMs to generate better documentation for obscure business logic, which makes code review AI agents more effective.

What always works - providing AI agents with more context via markdown files inside code repositories. You need to invest some effort upfront to write clear documentation that AI agents can read (into the context). These docs could be about your internal APIs, preferred architectural patterns, security constraints, etc.

Now, the prospect of writing docs makes many developers’ heads explode, but the good news is that Claude Code is great at generating these documents. Use it.

Unfortunately - in my experience - most enterprises really struggle with documenting even the most basic things about how they ship code. A lot of that knowledge sits in people’s heads (tribal knowledge).

That’s another reason why enterprises are bound to move slower with vibe coding, unless they fix cultural issues. Every CTO should focus on documenting tribal knowledge such that AI agents can move more effectively.

In terms of which vendor to use (Codex, Claude, Cursor, Devin, etc) - my advice is to just pick one and get started. My preference is just using Claude Code if you are a small team. For larger enterprises, maybe you need the collaboration features baked into Cursor or Devin.

Examples of documents to create and maintain to empower your code review (and coding agents):

• Internal style guides and architectural diagrams

• Your specific security requirements (e.g., “all database queries must use parameterized statements,” “all API calls must have rate limiting policy X attached”), etc.

Priority 2: Streamline the downstream processes in LOCKSTEP with vibe coding

You need to start investing in DevSecOps automation from Day 1.

Most companies want to adopt Cursor or Copilot at scale first, then figure out the rest. That’s only going to lead to piles of PRs and fatigue for the senior developers doing code reviews.

Instead, don’t scale code generation without streamlining security scanning, testing, and deployment. You need to manage the load at every step of the process, and unblock blockages.

Of course, since technology is moving fast, whatever you build may get obsolete in 3 months. But that shouldn’t deter you from getting started, since it’s an opportunity to learn.

Instead, iterate on your processes (Kaizen method), and unblock whatever nodes in the process that’s constrained. Start by automating more coding. Then, automate review, security, and testing. Then increase code generation velocity again. Then update your automated test suite. Rinse and repeat.

At the minimum:

1. Set up basic security scanning agents that runs on every PR, which already perform better than most human developers at flagging security issues. Both Anthropic and OpenAI offer Agents that perform extensive security reviews on your PRs. You can then layer on your company and domain specific policies on that. Start there and see how it performs instead of throwing your hands up.

2. Leverage agents to completely automate creating and execute test suites prior to creating PRs. Have agents periodically scan your code base to increase test coverage. You can literally just ask Claude to do just that.

3. Automate big chunks of browser based QA with LLMs. Yes, LLMs can use browsers and perform QA now. You shouldn’t be relying human QA team members for performing basic end to end tests.

4. Set up agentic workflows that assist with incident response so you can catch issues in production and triage faster. Agents are faster than most human developers at performing root cause analysis. Leverage that.

The pipeline needs to handle the volume before you increase the volume. If you need specific advice on how to create these automations, or which vendors to leverage, DM me.

Priority 3: Consider using managed services for things like security ops

There are some areas where it’s wiser to buy, versus building in-house solutions. One of them is all aspects of security (security review, threat detection, sec ops, etc).

One side effect of vibe coding is that it has made cyber security infinitely more challenging. Smart AI agents cause a paradigm shift in security because, unlike humans:

• agents can adapt in real time, autonomously, to the target’s security defenses

• agents are much cheaper to run than hiring human hackers

• agents don’t fatigue, they run 24/7

• it’s been known that agents are already becoming better at humans when it comes to hacking

Defending agains agentic cyber security threats is not something you should own. The area is evolving fast. You want these off your plate completely, not another thing to maintain.

That said, I don’t think traditional vendors like Crowdstrike or Okta are well equipped to provide solutions for this as well. They are too fragmented, solving different parts of the security problem (device management versus identity.. but they should all be integrated).

I would lean more on working directly with OpenAI and Anthropic for guidance on how to handle security reviews and cybersecurity.

If you are an institution or corporate, we offer group subscription (get 15% off for 3+ seats). I’ll remember you fondly, and you will build good karma.

Give a gift subscription

Quick Landscape: What’s Out There

This is not an endorsement of any company in particular, but a small subset of vendors you can check out.

Code review: Obviously the massive painpoint right now, and a lot of innovation is happening. CodeRabbit, Graphite exist, as well as Devin, SourceGraph, etc. Everyone’s putting out a code review offering, but none of them seems like the “clear winner”.

Security reviews, security ops: The labs (OpenAI, Anthropic) will bundle security review capability into coding agents, and I expect them to offer products as well. But specialized vendors can still add value. Snyk, Semgrep, etc, offer AI powered automated vulnerability detection. There’s also an overlap with the AI native cybersecurity market, which has companies like Abnormal (email security), Legion (sec ops), etc.

Testing/QA: Out of the box, coding agents are already really good at generating tests, and managing test suites. QA Wolf, Mabl, Testim, Mometic handle automated UI and regression testing. Less incumbent dominance here than other categories. More greenfield opportunity.

Observability and Incident Response: Datadog, Honeycomb dominate runtime monitoring and incident response - and they have AI agent features that automate parts of SRE tasks. Traversal is just one of many startups that are trying to sell a full blown AI agent in this market.

What’s the startup opportunity? Where should you build in DevTools?

Read more

• OpenAI’s ad business is misunderstood

• Claude Cowork Nukes Every Startup Playbook

Subscribe now

If you’ve been paying attention to the AI discourse lately, you’ve probably been exposed to some variations of these narratives:

• “OpenAI is losing the enterprise market to Anthropic” — somewhat true, as Anthropic’s been crushing both PLG (Claude Code) and top-down sales (AWS/GCP/Accenture partnerships)

• “Once LLMs are commoditized, all value will accrue to applications that have the ‘context graph’” — a view Marc Benioff (CEO, Salesforce) understandably loves

But those views are only correct if you naively assume that businesses are static, compute supply is infinite, and OpenAI doesn’t make any adjustments to its strategy. Unfortunately, none of these assumptions hold in real life.

In fact, there’s mounting evidence that OpenAI’s enterprise will regain momentum in H1 2026 based on a new revenue / engagement model.

To understand this, we need to cover a bit of recent history.

The second half of last year (2025) was about OpenAI setting the stage for its ad business and driving inference costs low enough to sustain a free tier, which is 4x the size of Gemini and 20x Anthropic. So OpenAI took foot off the gas from enterprise, while focusing on lowering its inference costs to prevent its Free Tier from bankrupting OpenAI.

That diversion created an opening for Anthropic to capitalize, and capitalize it did with style. Kudos to Anthropic.

But now that inference costs are under control, and the ad infrastructure is in place, OpenAI is revving up its enterprise motions. It started with OpenAI launching a post-sales / consulting arm, partnering with PE firms to sell AI transformation, and poaching Denise Dresser (former Slack CEO) as the new CRO to focus on enterprise.

Then a few days ago, Sarah Friar - OpenAI’s CFO - published a blog post titled “A business that scales with the value of intelligence” that reads like an investor update, but is actually a sneak peak at OpenAI’s revenue model for years to come.

In particular, she suggests that OpenAI’s revenue model will increasingly move from selling tokens through APIs, toward outcome-based arrangements where OpenAI shares revenue with its customers. In short, OpenAI will increasingly pursue value based pricing, taking a fraction % of value created, instead of selling raw tokens.

Why now? For a while, it’s been known that selling tokens via API is a crappy business, as it gives more bargaining power to application-layer customers, while enabling competitors to copy you. This necessitated a new engagement / business model, where OpenAI (or Anthropic) gets an explicit cut of the value for every dollar that every agentic application earns.

But you can’t just raise prices on tokens, because it motivates customers to seek out alternatives, such as open source models. So what’s the right strategy to capture more value, avoid selling raw API tokens, while keeping customers happy?

In this post, I’ll cover:

• OpenAI’s new engagement model with enterprise customers, which will maximize take rates and minimizes model provider churn

• Two additional levers that OpenAI can pull to maximize model revenue

• Why it will work

• Signals from analyzing more than 400+ job openings at OpenAI to reverse engineer its org structure

• Why the “foundation models are commoditizing” narrative may age poorly.

• What this means for investors and founders who are concerned about building something that OpenAI or Anthropic will make obsolete.

This is a founding / executive tier article. For executive tier members, we publish strategy deep dives from operator / investor lens. Prices will go up end of this month. See testimonials.

Paid tier readers will have access to my tactical market commentary focusing on the biggest shifts in AI (like this), as well as my paid post archive. Support this publication by joining paid tier. Join hundreds in our community.

Subscribe now

If you are an institution or corporate, we offer group subscription (get 15% off for 3+ seats).

Get 15% off a group subscription

Read more

• Claude Cowork Nukes Every Startup Playbook

• The Great Software Repricing

Subscribe now

Yesterday, OpenAI announced that they’re testing ads in ChatGPT’s Free and Go (the $8/mo tier) versions. Understandably, Twitter immediately lit up with negative reactions: “Sam’s desperation move,” “they got cooked by Claude,” “adding ads after slipping on enterprise”, etc.

Thus, the sentiment on OpenAI is squarely in the “negative” territory. And for good reasons. Claude Code is eating their lunch in developer and enterprise mindshare. ChatGPT growth stalled in Q4 2025 on a sequential basis. Its ventures into wearables and robotics make them seem scatterbrained and unfocused. Just like that, the narrative on OpenAI has shifted from “inevitable winner” to “desperate incumbent.”

So when the ads announcement dropped, it landed as a confirmation of the bear case.

But here’s what the dunking misses. First off, OpenAI always said they were going to be ad supported partially. I even wrote about why back in June 2024, that ad business is necessary for OpenAI as a hedge against LLM commoditization.

More importantly, OpenAI might have just announced the most important shift in direct response advertising in 20 years, since AdWords. They literally just invented a new ad format in a greenfield channel. And because everyone’s so bearish on them right now, nobody’s correctly pricing in what happens if it actually works.

When you focus just on facts and numbers, it’s unlikely that ads become OpenAI’s demise, but rather help it gain marketshare away from Google and Meta.

What Actually Happened

But first, quick level set on the launch.

OpenAI said it will test ads in “coming weeks” for free and Go tier (the $8/mo tier) users in the U.S. only. The format: sponsored products and services at the bottom of answers, clearly labeled, separate from the organic response. Pro, Business, and Enterprise tiers remain ad-free.

The principles they laid out:

• ads never influence answers,

• conversations stay private from advertisers,

• users can turn off personalization,

• they won’t show ads to under-18 users or near sensitive topics like health and politics.

The example they showed is revealing: a query about “best Mexican dinner recipes” returning a hot sauce product placement alongside the response (see below). That’s not accidental positioning. It ties well with OpenAI’s recent ambitions to grow its e-commerce partnerships revenue (w/ Walmart, Shopify, Target, etc).

You can imagine how these ads are useful for driving e-commerce revenue, but that’s barely scratching the surface.

Paid tier readers will have access to my tactical market commentary focusing on the biggest shifts in AI (like this), as well as my paid post archive. Support this publication by joining paid tier. Join hundreds in our community.

For executive tier members, we also publish strategy deep dives from operator / investor lens. Prices will go up end of this month. See testimonials.

Subscribe now

Read more

Software stocks have been underperforming broad indices (QQQ, SPY) and semis (SMH) since 2023 (guess what happened on November 2022). And recently, the selloff accelerated to start off 2026.

Many attribute the underperformance to seat compression worries, high interest rates, disappo…

Read more

Subscribe now

The incumbents (with distribution) are starting to fully embrace vibe coding. And startups that historically relied on quickness and smarts are NOT ready for what’s coming at them. A significant rethink in startup playbooks is in order.

Nine days ago, I published an essay arguing Claude Code needed a “Windows moment” - a GUI for non-developers. Basically, I called for someone to own the Desktop app opportunity for wrapping around Claude Code.

Speak of the devil.

Yesterday, Anthropic shipped Cowork, a macOS Desktop app that wraps Claude Code in a GUI. What’s notable: the app was fully vibe coded (not a single line of hand-written code), built in 10 days according to Claude Code creator Boris Cherny’s tweet.

Yes, the product (only available to Max plans) has rough edges, and isn’t exactly the vision that I laid out. But the significance shouldn’t be lost. A $350bn company with the biggest developer-facing distribution (after Microsoft) just shipped a vibe coded app to production, almost as fast as I could write a Substack piece.

Now, things get complicated when incumbents embrace the “vibeMaxx” movement, out of the fear that they, too, might get disrupted if they don’t ship faster. Most obviously, scaling startups will get harder.

Unfortunately, it’s already happening. Incumbent SaaS feature velocity is skyrocketing:

• Salesforce pivoted AgentForce in 6 months after getting panned, and they just launched an Anthropic-backed Slackbot in 3 months since announcement.

• Notion revamped Notion Agents in 6 months.

• Hubspot released a V1 of their AI agent for growth in a quarter.

Note, these datapoints are from BEFORE the recent agentic AI step-up with Claude Code v2 in December 2025.

Things are only accelerating. The incumbents are working harder, cultures are shifting to embrace YOLO, and it’s only a matter of time before nimbleness, incentives, and informational asymmetry—which used to favor startups—evaporate thanks to agentic coding. Not that it will save the incumbents (CRM down 7% today even after releasing Agents for Slack) - they will, however, swim a lot faster.

But, here’s the uncomfortable truth: vibe coding was supposed to level the playing field. Instead, it’s poised to amplify existing distribution advantages. When incumbents move at startup speed, they absorb 1→N innovation faster than startups can reach market. Funnily enough, I said the exact same thing 18 months ago in “Why Innovation’s Dilemma Doesn’t Apply to AI”.

Of course, as Anthropic trailblazes the agentic AI category, there may be greater output from startups building different variations of Claude Cowork, as TAM increases. That argument, however, drastically underestimates how “captureable” TAM shrinks for a single given startup (point estimate). There are multiple historical precedents of this, which I’ll get into.

But here’s the kicker. The takeaway isn’t “Distribution is the New Moat” (another popular LinkedIn thought-slop I debunked last year). Rather, distribution is just the price of a lottery ticket, not the lottery itself.

That said, not all hope is lost. So in this essay, I’ll offer thoughts on how to re-calibrate against the new reality and discuss the industry-wide implications. And you must do that now, before spending a year competing in a hopeless market, only to be disrupted by the next Claude release and you are stuck running a business you hate.

What Just Happened

But first, let’s catch up on the Claude Cowork news.

In my Claude Code is MS-DOS piece, I said Claude Code needed a GUI—something that preserved its file system power but worked for non-developers.

Claude Cowork is just that. You can drop any folder inside your desktop, and ask it to do some work (e.g. create a PRD), and it does it.

But it does something even smarter than Claude Code. It seems able to generate and leverage Skills on the fly to accomplish tasks. For example, I gave it a prompt to generate a PDF, and it dynamically pulled a PDF processing Skill to assist itself.

Signs of continual learning, if you ask me.

Clearly, it’s not far from a world where Claude Cowork actively distills user preferences into Skills, such that end users don’t have to think about these low-level constructs.

But the real significance is in how quickly the Claude Code team put this together. Yes, you may be skeptical about their claims of 100% vibe coding it. But having vibe coded large apps myself, I think they can be trusted.

The product details matter less than the timeline. 1.5 weeks. That’s the new normal for what a well-resourced team with vibe coding can ship.

Now, let’s talk strategy and implications.

I recently launched executive-tier (founding tier) to help operators, VCs, and founders specifically with decision making. Executive tier subscribers also get access to my Predictions for 2026 in AI and company level deep dives in enterprise AI.

Prices for executive tier will increase in end of this month by 20%.

My writing has already helped hundreds of investors/execs make the right decisions. See Testimonials for Enterprise AI Trends.

Subscribe now

**Institutional investors must subscribe via group subscriptions**

Get 15% off a group subscription

Why Velocity Parity Breaks Everything

Read more

I’m writing this post in part to challenge that view. When technology changes very rapidly, things like enterprise inertia may not protect a software business in the long run (~2 years).

Instead of debating hypotheticals and vibe posting about “moats” on Twitter, I decided to run an experiment.

Using Claude Code (btw, I’m not shilling for anyone, Codex or Gemini CLI works just fine too), I vibe coded a barebones Salesforce clone over three hours.

Then afterwards, I had Claude Code reflexively use Chrome to go through every logged in page in the CRM, and pull out all the data inside it. It was able to do it autonomously in one shot.

So what does this mean? Am I saying all software companies are doomed, and you should start doing Plumbing / HVAC / Botox businesses? Or start an “AI consultancy”? Not quite. Don’t do that to yourself.

Note, I’m familiar with Salesforce as a former user (at AWS) and have written multiple constructive articles about it (see my coverage on AgentForce, and coverage of Salesforce’s AELAS for exec tier subscribers). I’m also well aware of the bull/bear arguments in this space, and have traded SaaS for ten years.

With that in context, in the rest of this post, I’ll share more on my experiment, and what I make of software moats as of Jan 2026.

I recently launched executive-tier (founding tier) to help operators, VCs, and founders specifically with decision making. Executive tier subscribers also get access to my Predictions for 2026 in AI and company level deep dives in enterprise AI.

Prices for executive tier will increase in end of this month by 20%.

My writing has already helped hundreds of investors/execs make the right decisions. See Testimonials for Enterprise AI Trends.

Subscribe now

**Institutional investors must subscribe via group subscriptions**

Get 15% off a group subscription

How this actually works

Read more