What I’m thinking about: How fast AI keeps getting cheaper and more capable… and where the real edge (the oldest fundamental there is) actually comes from.
The state-of-the-art Claude Opus 4.8 model dropped a few weeks back, and it came with more controls for squeezing throughput out of every token (“state-of-the-art” until Fable 5 dropped earlier this week…only to be pulled this weekend due to a US government directive, man this industry moves fast).
For most work that’s well suited to AI (“left-brain” output, rooted in logic and mathematics), I’m firmly in the camp that the machine now out-produces the marginal human hire.
Our own output keeps compounding week over week, and it’s running on a double-factored gain… the models keep getting better AND our skill at wielding them improves at the same time.
It doesn’t always show up in an immediate bottom line (in our industry, you still have to account for a risk-off market and real estate that’s moving slower than it has in decades).
But it’s unmistakable that my team and I are producing WAY more than we were a year ago, let alone five years.
So when I see the metrics floating around claiming AI is basically break-even against human talent (sometimes even a touch worse), I want the context behind them.
I want to know how well-trained the people using the tools actually are, what outputs and KPIs they’re tracking, and whether they’re on subsidized plans or paying for raw API tokens (the latter gets expensive fast).
The US frontier models (think OpenAI, Anthropic, Gemini) cost real money. The open-source Chinese LLMs run 10 to 30x cheaper…but they aren’t frontier-grade, so you’re trading away the top tier of intelligence to get there (I’m generally of the opinion that accessing the highest grade of intelligence is always worth the cost, outside of routine, bracketed tasks with known gold standard outputs).
Why Salesforce is glad to burn $300M on Claude
Two things are happening under the hood.
First, a lot of companies are genuinely wasteful with how they spend tokens (look up ‘token-maxxing’).
But even when the usage runs hot in the near term, getting to the answer faster (or serving your market faster) is often the better trade.
The operators experimenting (targeting the core biz constraint, of course) most aggressively right now are quietly building the widest moat, while everyone else waits for the cost to come down (or even worse, still keeping practical AI applications on the back burner).
Second, and bigger: Marc Benioff said on the All-In podcast that Salesforce will spend close to $300M on Anthropic tokens this year, almost all of it on coding. For a company that size (~$136B market cap, ~$15B in annual free cash flow), that figure might even land on the modest end.
The number grabbed the headlines. What stuck with me was his read on where this all goes.
Right now it’s messy how the AI models hand out tokens… every user and every team effectively starting from scratch, mountains of redundant work getting rebuilt over and over, disgusting amounts of waste.
(Real estate analogy: imagine a master-planned community where each residence had their own water and sewer lines, individually piping back to the central reservoir, but no shared throughput in the neighborhood. An engineering nightmare and costs would absolutely skyrocket.)
Benioff sees an intermediary layer emerging, something that sits between the base AI models (e.g. Claude or ChatGPT) and the end user, and routes the unique reasoning to an LLM (where token usage is justified) while pulling already-proven, existing frameworks (e.g. a validated mortgage calculator) from public or private cloud libraries.
We have watched this exact movie before in pre-AI software engineering. Once somebody builds a reusable framework (e.g. Ruby on Rails), nobody rebuilds it from scratch again.
As another analogy, picture a front door…if the same door fits a hundred different houses, you cut that wood ONCE and just swap the aesthetics on top.
Squarespace and WordPress did the same thing with web templates: thousands of pre-built frames so nobody starts from zero.
That is coming for AI builds too. A hundred million people asking for the same style of keynote deck should not burn a hundred million separate token bills rebuilding it from nothing.
Hearing one of the sharpest minds in tech lay this out was a humbling moment, frankly, and it was a reminder that there are always additional levels to the game, and in order to compete at the top, you have to hone your capability to see further than anyone else.
The shift toward AI frameworks seems near-certain to me (likely there are already shops working on building that intermediary layer right now)…the only open question is timing, and I wouldn’t be surprised if we start to see those before the end of the year.
The vibe-coding wave runs straight into a wall
All of this relates to a question plenty of people keep asking out loud: whether software is even investable anymore.
Long term, that is genuinely TBD.
But my read is that the big enterprise players (e.g. Microsoft’s Office stack, Salesforce) win first, because they are pouring hundreds of millions (perhaps billions) of dollars into AI to sharpen products that are already dominant.
The flip side is the wave of smaller products getting spun up on tools like Replit or Lovable. The underlying AI models, as good as they are, WON’T solve those products’ bandwidth problem.
I’m serving as an AI advisor for Callan Faulkner’s Uncommon Business now, and I watched this play out in real time on their recent Accelerate to Automate (A2A) launch.
They pulled in FAR more customers than they had planned for, and the plumbing buckled…Zapier, the automations adding events to client calendars, Zoom, all of it broke under the load.
They had to upgrade to enterprise-level plans (and/or new software) in real time to service their clients (FYI it only took a few thousand people to break the existing systems, not millions).
The lesson sits there in plain sight: you can spin up clever software in a weekend, but you cannot (yet) spin up the rails to carry real volume (that’s where enterprise SaaS still shines).
=====
In the midst of market turmoil and questions about the continued viability of SaaS, I love how Benioff hand-waved that away to bring the conversation back to what companies can actually control: whether you deliver more value to your customer than the next operator does.
The foundation of ANY company is its core offer. (There’s a reason Hormozi wrote $100M Offers first…if you don’t have an offer, everything else is moot.)
My mentor Ryan Levesque sharpened the same point with a Venn diagram he recently illustrated for his upcoming book, Return to Real.
Whatever core problem you are solving for yourself almost always overlaps with what your market needs solved… so solve your own problem well enough, and more often than not you have already found your market.
That is the durable truth sitting under all the AI noise. The models will keep changing on a clock none of us sets, and serving the customer better than anyone else will keep being the thing that endures…forever.
How this is already reshaping what we build
This is the exact lens I’m using to build the Cowork plugin for SLC Chat right now.
The feedback, internal and external, keeps landing in the same place: it has to be faster, more automated, able to pull comps through whatever pathway gets there quickest (maybe a Claude Code setup instead of Claude Cowork… the interface matters far less than the speed and reliability).
This last week was particularly deal-heavy (perhaps our busiest in 2026, and funding docs being prepped as I write this), and all I kept thinking about was that I wanted to drop these deals into a system that goes to work on its own, pulls the right comps to start from, and lets my team and me make sharper calls from there.
The SLC Chat web app was a strong first test case (it taught us how the API tokens actually behave, and what genuinely serves an operator versus what just looks impressive), and we are pivoting toward a Cowork plugin build because it serves our own needs AND our audience’s better at the same time.
That is the building-blocks idea pointed straight back at our own shop: the more reusable underwriting infrastructure we create, the less our potential operator partners (ourselves included) have to rebuild from scratch on every single deal.
=====
Our underwriting process (and access to capital) is still our core offer…AI is just the leverage we are using to deliver it.
We write checks from $50K+. We close 100% of the deals we commit to. And we bring national underwriting built across thousands of reviewed transactions, levered by the most intelligent AI models in the world.
