# AI capex — X 热门讨论 (2026-09-28 01:06 UTC)
## @SMB_Attorney (SMB Attorney) · 09-28 00:06 · ♥79 ↻5 💬21 “We’re going to buy businesses and implement AI.”
Guys, with the exception of maybe a handful of people or firms, I’m not sure many people have a better front-row seat to the small and lower-middle-market acquisition ecosystem than we do.
We take 200+ calls every month from prospective buyers, searchers, independent sponsors, family offices and private equity funds.
And that doesn’t include the orders of magnitude more conversations we see through communities, group chats, conferences and the hundreds of transactions that move through our firm.
Based on our own deal volume and what we see in the market, we believe we’re among the most active law firms serving the lower-middle-market M&A ecosystem in the country.
And I can tell you:
“We’re going to buy businesses and implement AI” is EVERYONE’S thesis right now.
Greg’s piece is thought-provoking, and I think his broader point is right: everyone buying or operating a business should be thinking seriously about how AI can be layered into the company to improve productivity, reduce costs and expand margins.
That is absolutely where things are going.
And damn it, it’s a GOOD thesis.
I say “good,” not “great,” because there is a huge assumption buried inside it:
That you actually get to the part where you can focus on growth and AI implementation.
The acquisition lifecycle usually looks something like:
1. Acquire 2. Stabilize 3. Grow 4. Exit
AI implementation is largely part of Step 3... Growth.
But before you get to Step 3, you have to survive Step 2.
And Step 2 is where things get real.
> You inherit employees you didn’t hire. > Customers you didn’t acquire. > Vendors you didn’t negotiate with. > Processes you didn’t design. > Technology you didn’t choose. > Financial statements you didn’t prepare. > A culture you didn’t create.
And often years of institutional knowledge sitting inside the head of a seller who just got wired a life-changing amount of money and is mentally halfway to the beach.
Meanwhile, the debt payment is due every month.
A lot of acquirers get stuck somewhere between Stabilize and Grow for years.
> They fix people problems. > They replace customers. > They rebuild accounting. > They improve working capital. > They learn the industry.
They deal with equipment breaking, key employees quitting, sellers behaving strangely, unexpected capex and a thousand other things that never appeared in the CIM.
They put out near CONSTANT fires.
They service the debt.
They hold on for dear life.
They may still create seven figures of wealth...
But they never really reach the clean, optimized growth phase they imagined when they built the model.
That’s why I think assuming you can simply acquire a business and quickly increase EBITDA margins by the magnitude Greg is describing will be wishful thinking in most cases.
Not because he’s wrong about the opportunity.
He isn’t.
AI can absolutely improve margins.
It can reduce administrative labor, improve sales processes, accelerate quoting, enhance customer service, automate workflows and create enormous operating leverage.
Everyone should be thinking about how to use it.
But growth requires stability, management bandwidth, good systems, clean data and capital.
AI is a tool.
It is not a substitute for competent operations.
You still have to buy the right company.
At the right price.
With the right capital structure.
You still have to retain the right people.
Protect the customer base.
Understand working capital.
Manage cash.
Navigate the transition.
And actually operate the thing.
So if you’re considering acquiring a lower-middle-market business, do your homework, hire good advisors and go in with your eyes wide open.
Greg is right that AI creates a massive opportunity for business owners and acquirers.
But the hard part is still getting the business into a position where you can actually capture it.
This is hard as FUCK.
And AI didn’t make that part any easier. > 引用 @gregisenberg: $5T opportunity: AI Roll Ups https://x.com/SMB_Attorney/status/2104362046372360469
## @MAGAMAHACindy (Cindy K) · 09-27 23:47 · ♥56 ↻6 💬4 The economy hit the gas.
Treasury Sec. Scott Bessent says the U.S. has entered “the acceleration phase.” Translation: the factory-building, AI-spending, tax-cut setup period is over.
Now the actual output, jobs, and productivity are starting to show up. CapEx is turning into production.
Blueprints are becoming paychecks. Growth is picking up speed. 📈 https://x.com/MAGAMAHACindy/status/2104357290249699342
## @MilkRoadAI (Milk Road AI) · 09-27 20:00 · ♥31 ↻7 💬4 The AI buildout just got $750 billion bigger in less than nine months.
At the end of 2025, analysts expected Google, Microsoft, Amazon, Meta, and Oracle to spend a combined $1.14 trillion across 2026 and 2027.
Less than nine months later, that estimate has jumped 66% to nearly $1.9 trillion and that does not mean the companies have already spent this money.
It means Wall Street has dramatically raised its expectations for how much they will invest over those two years.
Google and Amazon alone are now expected to spend more than $1 trillion, while Microsoft’s estimate has risen to $373 billion, Meta’s to $333 billion, and Oracle’s to $174 billion.
Most of that money will fund the physical infrastructure behind AI GPUs and custom chips, servers, memory, networking, data centers, power, and cooling.
The biggest takeaway is that the AI infrastructure boom is not slowing down but rather accelerating far faster than analysts expected.
Hyperscalers are effectively telling the market that compute capacity remains a competitive advantage and that underinvesting could be more dangerous than overspending.
That is extremely bullish for the companies selling the picks and shovels, but it also raises the stakes.
Goldman estimates these hyperscalers may need roughly $300 billion in annual AI revenue to justify the investment, meaning the next phase of the story will be about whether AI monetization can catch up with this historic spending spree.
I’m already positioned across the companies benefiting from this massive AI capex revision, from memory and networking to power, cooling, and AI cloud infrastructure.
If you want to see exactly what I hold to capture that spending, check out my full Milk Road Pro portfolio below.
https://t.co/thIhK9ZH4E https://x.com/MilkRoadAI/status/2104300061182324963
## @TejaswiPalam (Tejaswi | AI Industrial Shift 🇮🇳) · 09-27 13:39 · ♥30 ↻2 💬4 A ₹200 Cr company quietly moving into a different league
Simplex Castings Ltd
I looked at what changed
FY24 to FY26
Revenue: ₹122 Cr to ₹202 Cr ROCE: 10% to 24% Debt: ₹72 Cr to ₹50 Cr
Now TTM revenue is ₹219 Cr Q1 FY27 revenue hit ₹61 Cr Operating margin: 19%
But interesting part isn’t just numbers
Simplex makes heavy engineering castings, but also does fabrication, forging, machining and assembly
Now now customer mix is getting interesting
BHEL SAIL JSW Vedanta Mazagon Dock
Company also adding capacity
FY27 capex is planned at around ₹25 Cr
Management is targeting ₹300 Cr revenue in FY27 and ₹500 Cr over the next few years
That target still needs execution
Capital remains biggest thing I’m watching
But this is no longer just a small casting company
Now question is
Can Simplex turn a ₹200 Cr foundry into a ₹500 Cr integrated engineering business?
That’s the story I’m watching 👀 https://x.com/TejaswiPalam/status/2104204203392696559
## @JonkooTrades (Jonkoo Capital) · 09-26 23:45 · ♥33 ↻0 💬1 $NBIS - Nebius is increasingly becoming a bet on the conversion of power into revenue, while short-duration contracts give it unusual exposure to rising AI compute prices.
Nebius might have one of the more interesting setups in AI infrastructure right now, because the story is moving beyond simply “they have GPUs.” The real question is how quickly Nebius can turn contracted power into active compute, and then turn that compute into revenue.
Truist estimates Nebius can go from roughly 170 MW of active power in 2025E to 750 MW in 2026E, while ARR potentially jumps from around $1.25B to $7-9B.
That is an enormous conversion year. The infrastructure is being secured today, but a lot of the revenue associated with it has not actually hit the income statement yet.
And demand visibility is already becoming pretty significant. Truist estimates Nebius has around $54.5B of total contracted backlog, consisting of roughly $37.5B of reported RPO plus another ~$17B of incremental potential backlog.
The important part here isn't just the size of that number, but what happens as the capacity behind those contracts comes online. Contracted power has to become connected power, connected power becomes active power, and only then does it really start producing ARR.
That creates a pretty visible pipeline where infrastructure execution becomes the main bottleneck rather than finding customers.
But I think the more interesting part of the Nebius thesis is pricing. Consensus apparently assumes ARR per MW falls from roughly $10.7M in 2026E to $8.9M in 2027E. Yet Truist points out that recent 1-3 year contracts are being priced around $20-25M per MW, while short-duration 3-6 month deals can reach $40-50M per MW.
Nebius has historically used much shorter contracts than something like CoreWeave, meaning it doesn't lock all of its future capacity into today's economics. In a market where GPU availability remains tight and compute pricing is moving higher, that flexibility could become extremely valuable.
This also ties directly into the GPU price increases we've been watching from Nebius. If compute pricing continues moving higher, Nebius doesn't necessarily have to wait years for its contract book to reset. It can monetize part of its capacity through shorter-duration agreements, capture the higher spot economics, and still use large long-term contracts to finance the buildout.
Truist estimates roughly 70% of Nebius' 2026 deals include upfront prepayments covering around 50-60% of associated capex. So the Microsoft/Meta-style contracts provide financing and demand visibility, while shorter contracts potentially provide the margin and pricing upside. That's a much more interesting combination than simply locking every MW into a five-year contract.
The numbers get pretty wild if execution actually follows the capacity roadmap. Truist models roughly $3.35B of 2026 revenue and $12.0B in 2027, with 2027 ARR around $15B.
They also see Nebius eventually reaching 20-30% adjusted EBIT margins as utilization rises and the fixed infrastructure base gets absorbed by a much larger revenue stream. There is obviously a huge amount of capex between here and there, and depreciation matters enormously for a business constantly buying newer GPUs, but that is exactly why the next 12-18 months matter so much.
We are about to find out what the economics of this infrastructure look like once hundreds of additional MW actually become revenue-generating.
And despite the premium people often attach to Nebius, Truist's peer table makes the valuation more interesting than it initially looks. Their numbers put Nebius at roughly 5.9x 2027E EV/Sales versus a peer average around 5.1x, but with modeled revenue growth of roughly 357%.
So you're paying somewhat more than the group, but you're also underwriting a radically faster revenue ramp.
For me, that makes the biggest thing to watch from here pretty simple: active MW, ARR/MW and contract pricing. https://x.com/JonkooTrades/status/2103994276136820825