The next AI giants will be industrial companies, not software companies.
That’s because the biggest constraints in the AI revolution are becoming physical:
• Power
• Cooling
• Land
• Water
• Grid access
• Fiber
• Permits
• Transformers
• Substations
This is why data centers are becoming one of the most important parts of the AI story. (And why we're already talking about putting them in space.)
Data centers are no longer just buildings full of servers. They are factories that turn electricity into intelligence.
Training will still push toward massive centralized campuses. But inference is different.
As AI moves into cars, robots, hospitals, factories, defense systems, finance, and everyday business operations, compute will need to be closer to where it's being deployed.
That means the AI infrastructure map gets much more complicated.
• Massive liquid-cooled campuses
• Nuclear-adjacent facilities
• Edge compute everywhere
• Data centers under water or in space
The next phase of the revolution will be decided by the companies that understand power, physics, and infrastructure better than everyone else.
And as for humans, maybe our last job will be building and maintaining these places while the machines do all the thinking.
Joe Mayberry makes the case that the agentic era is here, but the industrial revolution for AI hasn't started. We've got the foundational models but still lack the connective tissue that turns isolated tools into something that scales.
Check out his full interview on Applied Intelligence.
OpenRouter just raised $113M.
The company gives developers access to 400+ AI models through one API, and says usage has grown 5x in six months to 25 trillion tokens per week.
That is the part people should pay attention to.
The future is not one model for everything.
We recently gave everyone at my company paid access to Claude so the whole team could experiment, build habits, and learn what AI can actually do inside the business.
But I also knew we would eventually outgrow a single-model approach. The more you plug into one system, the higher the switching costs get.
At the same time, the orchestration layer keeps getting better.
Companies like OpenRouter make it obvious where this is headed. A central place to work, with different models routed to different tasks based on quality, speed, and cost.
Our creative team is already seeing the need for a mid-market product like this. Claude is great, but the cost keeps growing, and we still have needs for other models where Claude falls short.
Using one LLM for everything will start to feel inefficient. It already does.
Keep that in mind as you roll out AI at your company.
SAP is worth $250B because it became the source of truth for enterprise operations.
The next $250B company will be worth much more because it made SAP a dumb database.
At Voi we replaced a lot of mid and long tail SaaS with custom internal software. Scheduling, ops dashboards, reconciliation layers. It worked.
At Klarna, some of the same people went further and went after the critical systems themselves. Replacing an ERP directly is brutally hard. It did not fully work at the core system level, but they understood enterprise software internals at a depth nobody else has.
The team that did that work is now building @pitdotcom together.
The real value was never inside the systems of record. It was in the human layer around them.
The person copying data from SAP into a spreadsheet. The analyst reconciling NetSuite against a supplier PDF. The ops manager chasing approvals by email because the workflow lives between systems and nobody connected them.
Real work, and the software just could not do it.
Until AI.
You build net new software that runs those workflows end to end. SAP stays. It just stops being where the work happens.
Phase 2 is more structural. As your software performs the work around a system of record, you extract its logic. You learn what it actually does in practice, not in theory. At that point it goes dumb. A database you act on via API. The value moves up the stack.
Phase 3 is the one nobody has built yet. The company that runs the execution layer across thousands of enterprises in the same vertical understands how those operations actually run at a depth no single enterprise can.
Every manufacturer sees its own workflows. The execution layer sees the patterns across all of them.
That asymmetry grows with every customer and every month in production. The systems of record captured the data. The execution layer captures the intelligence.
SAP spent 50 years becoming indispensable. The company that wins the next 50 is not building a better SAP.
It is building the layer where the work actually happens, compounding in ways SAP never could.
Depending on where you work, your next boss might just be AI.
Joe Mayberry, Head of AI at SailPoint, breaks down the "agentic epoch" and explains how we are rapidly moving toward a future where AI software agents are actively asking human employees to accomplish tasks.
Check out the full episode wherever you get your podcasts.
Shadow AI refers to employees using unapproved AI tools at work. And the bigger your shadow AI problem is, the worse your AI rollout probably was.
Microsoft says 78% of AI users are bringing their own tools to work. Translation: your team is already using AI. The only question is whether you know about it.
Most companies respond with warnings, permissions, and a 14-page policy doc no one reads.
That will not work.
People follow the path of least resistance, especially when that path helps them do better work faster. And more importantly, you should want your team to do better work faster.
You don’t want your customer data, contracts, code, financials, and strategy docs getting pasted into random tools with no visibility. Provide a better option.
If you want to fight shadow AI and win, make the approved path easier than the unapproved one.
Give employees AI subscriptions. Encourage usage. Provide training.
Make the safest path the easiest one.
Every company has 2 choices right now with AI:
• Keep debating internally on the best way to roll it out
• Accept you’ll probably overpay and make mistakes at first, but start anyway
If you’re ready to start, start here:
1/ Outcomes first
Focus on the results your team is accountable for, not the tools or processes you currently use. Things like: evaluating expenses by vendor, identifying winning creative, improving reporting, writing more effective scripts, etc.
2/ Identify the constraint
Ask what work currently slows your team down. Things like: research and information gathering, manual data analysis, building reports, etc.
3/ Ask the “10x question”
Imagine what would happen if a key constraint disappeared. Ask yourself: what would this outcome look like if it were 10x better?
If the answer involves manual work, slow analysis, or fragmented data, there is likely an AI opportunity worth exploring.
4/ Imagine a "smart intern"
What if you had a brilliant intern who could read everything, write anything, analyze data instantly, and never sleep? What work would you give them?
5/ Focus on decision bottlenecks
The biggest constraint is often a hard decision to make because the right information isn't readily available. Consider the most important decisions your team makes regularly, then ask: what information would make those decisions dramatically better?
A major media CEO just told his team to budget for zero search traffic in the coming years.
In the latest episode of Applied Intelligence, @mcia_ explains why publishers are in trouble while brands have a massive new opportunity.
AI tokens are about to become a utility bill. Except unlike water or gas, usage is still wildly unpredictable. And as intelligence becomes a commodity, the companies that truly understand and optimize utilization will have a massive advantage.
Naturally, there are tons of companies selling solutions to this problem to Fortune 1000.
Token forecasting. Token budgeting. Model routing.
At that scale, you’re guaranteed to have AI spend leakage. But the more interesting question is what happens to everyone else.
What is the best way for a small business to know if they are better off hiring a $150,000 business intelligence analyst or spinning up an agent?
Will the agent cost $5,000 a year? $50,000? $500,000?
Right now, most small businesses are piecing this together from fragments.
If you use Perplexity, you can go back through chats and at least see credits used, and Claude gives you a general sense of usage. Other tools might tell you who used tokens and when.
But none of that really tells you whether the spend produced anything valuable. You can have daily meetings and review outputs, so you are not flying blind. But you are still mostly guessing.
How much compute will this project actually take? Which AI projects are worth pursuing? Which employees are creating meaningful output, and which ones are just using up tokens?
More AI tools need to build this right into the product to give small businesses a chance at being competitive.
Because when everyone has access to frontier intelligence, how efficiently you deploy it becomes your moat.
If you thought your most annoying board member was bad before AI, strap in.
There have always been 2 types of annoying board members:
The one who clearly read none of your materials, shows up cold, and burns the first 30 minutes asking questions that were answered on page 3.
And the one who read everything, wants everyone to know it, and cites the third footnote under the table on page 57.
AI will make both of them worse.
Before, someone had to prepare the 200-page board deck, and directors had to actually read it.
Now, the underprepared director can have AI analyze the presentation and spit out dozens of “smart sounding questions” so it looks like they know what they’re talking about — thus becoming 10x more annoying.
And the overprepared director, who actually did read the deck, can now use AI to run even deeper analysis, challenge every point made in the presentation, and be 10x more annoying, showing everyone how well prepared he is.
And they are going to do this all while asking about your AI strategy. AI doesn’t change behavior. It magnifies it.
Talk to your board about the risks of AI making members more annoying.
Are legacy software moats dead? @Samsara VP of Engineering @lefthandmagic explains why AI-driven computer use is rendering traditional APIs and complex integrations obsolete.
How @Samsara speeds up engineering with a fail-fast AI mindset. VP of Engineering Praveen Murugesan ( @lefthandmagic ) shares how AI tools are enabling engineers to build prototypes instantly, tackle bugs on the fly, and run multiple AI agents across virtual desktops simultaneously.
We gave everyone at my company access to Claude. Then we had to make a rule.
Just because Claude can make a 30-page presentation or research report, does not mean you should share out all 30 pages!
This sounds obvious until you see what happens inside a company.
Someone who never would have written a 30-page market analysis doc now has one ready in five minutes. Then the recipient asks Claude to summarize it into four bullets. Then someone else asks Claude to turn those four bullets back into a strategy memo.
At that point, no one is actually reading the work. One AI is creating the document. Another AI is summarizing it. Another AI is expanding it again.
Claude can be a great co-thinker and co-partner, but boy does he tend to be wordy. So, brevity and context setting matter more than ever:
• Is this important for me to understand, or important for everyone to understand?
• Does the whole team need the background, or do they just need the decision?
• What does someone actually need to know to do their job better?
When anyone can create decks, memos, and presentations almost instantly, wasted time becomes a much bigger risk. The output looks productive, but the cost gets pushed onto everyone else.
This has been one of many learnings as our team experiments with AI: It makes it easier to produce more work, but it also makes it easier to create more work for other people.
Curious if anyone else has had interesting learnings, rules, or norms come out of their team using AI?
@gavinpurcell@AIForHumansShow I think it's funny how you try to suck up in public but I have on good authority that you and @Attack i've been calling Claude and controlling monster in private
One week ago, we gave 100 people at my company access to paid plans on Claude. Here are the most interesting wins in week one:
• Finance pulled vendor-level details from every credit card transaction in NetSuite. Weeks of work, now minutes. Already building the list of SaaS to cut.
• Our eCommerce team built a TikTok Shop outreach skill. 24 hours later: $162 in GMV and 7 lapsed customers back.
• Our tech team built two always-live dashboards that auto-flag conversion drops and channel anomalies. Replaced a daily manual routine.
• Our Amazon team rebuilt our PPC bidding rules from the ground up. Closed gaps that were bleeding spend.
• Our tech team built a robot shopper that visits all 13 of our Shopify storefronts daily, tests checkout, and Slacks plain-English alerts when something breaks.
• The creative team cut the effort involved in B-roll mapping from hours to 15 minutes. For every single video.
• Daily Seller Central monitoring across every brand now runs itself. Account health, hijackers, stranded inventory, suppressed listings. All auto-posts to Slack.
• An analyst who's been chipping away at an inventory recovery problem since 2022 finalized the system architecture this week.
The whole company is doing this. Finance. Creative. Merchandising. eCommerce.
We have no master plan or AI strategy deck. We’ll build that plan based off of the real-word wins we see from experimentation.
If you’ve ever said “We’d have to cut other tools before investing in AI experimentation for the team,” realize that my finance team is building a full list now of tools we’re canceling next week, based on results from our experiments.
Find the money to let your team experiment with AI.
As I’ve become more of a manager and less of a doer throughout my career, I’ve worried about losing “actual” skills: building good models, making good presentations, etc. I thought without them, I risked being irrelevant.
I eventually realized my ability to do manual work in Excel was never the differentiator. It was my ability to think strategically and get the best out of those I worked with.
I was reflecting on this recently while using Claude for Excel, and realized that we’re all already irrelevant by the old definition.
With Claude, I built a great model without doing anything. I wrote no formulas. I just told it the outcome I wanted:
• “Ad spend versus profit, last year versus this year”
• “Fastest growing expense category”
• “Our ROAS versus industry benchmarks”
For everyone now, being good at your job is no longer about knowing how to do the work. It’s not the “actual” skills.
It is about knowing what work should be done and getting the most out of your AI colleague.
Education and entry-level work evolved to make sense in a world where everyone had a computer in their pocket and access to every answer.
Now they have to do it again to make sense in a world where technical skills and know-how have been completely democratized.
If you’re getting your AI playbook just from Twitter, you’re wasting your time.
“I built an agent in 10 minutes!”
“This ONE prompt changed everything!”
“This AI hack makes me $10k/month!”
None of this is real. These are people selling you a story to drive algorithmic engagement or to sell you a service, course, or who knows what else. They are the modern-day version of the ThighMaster infomercials: Lose weight in 10 minutes with no work!
You can drive real results with AI, but there is no magic prompt or instant agents. You need to put real work into understanding the outcome and approach, and then you need to iterate.
These sensationalist posts flood our feeds because algorithms show us what feels easy and exciting, because it gives us dopamine.
But dopamine doesn’t get you results. It starts you down new paths toward shiny objects.
From A to B, but never to C.
I’ve had conversations with 100+ operators and PE firms over the past 12 months, and I can tell you with certainty that none of them have experienced meaningful leaps forward:
• From an agent built in 10 minutes
• From a single prompt
• Or from any AI hack
We gave our entire team access to Claude. $25 or $125/month per person. 100 employees. It will end up being north of $50,000 of spend with usage overages.
It will be the highest ROI decision we make all year.
The goal: experimentation.
We gave everyone access and told them to:
• Try new things with Claude every day
• Start with desired outcomes
• Figure out how to get there 10x faster with AI
• Share wins in weekly town hall meetings and Slack
That’s it. No master plan. No “AI strategy” deck.
Just reps.
What’s surprising is how many companies are still choosing not to do this.
I’ve heard:
• “We want software spend to go down, not up…”
• “We’d have to cut other tools before doing that…”
• “We don’t have an R&D budget for AI experimentation…”
Meanwhile, the companies that are experimenting with AI and getting good at using it are not just saving on software spend, they’re growing revenue faster and leaner.
Trust me. Find the money.