There Is No AI-Insulated Industry

I round up the most relevant AI-in-finance news, the deals being done, who's rolling out what, and what's actually working on the front lines.

Wall Street's largest legal buyers have started telling law firms that AI savings belong partly to the client. Morgan Stanley and Citi want new fee arrangements, and Goldman Sachs is asking how much faster the work has become.

Nvidia has agreed to buy Hugging Face for roughly $13 billion, and Anthropic is close to handing Morgan Stanley and Goldman Sachs the top roles on a $2 trillion IPO.

Elsewhere: private credit has marked down its software loans again, Microsoft has reorganised around agents and infrastructure, and OpenAI has released Astra.

But first, my failed attempt to name an industry insulated from AI, and why the search ends at the balance sheet.

In today's Acquisition Intelligence:

From The Trenches

There Is No AI-Insulated Industry

Someone asked me recently which industries I thought were most insulated from AI. It is a question that has been on a lot of investors' lips for the past 24 months, and the prevailing answer runs something like this: anything software-based is at risk, knowledge-work businesses are exposed but can try to capture the gains themselves, and physical infrastructure is immune, at least for the next five to ten years.

The more I thought about it, the more I struggled to name one. There are no AI-insulated industries any more. Everything gets disrupted, and the only question is when.

That on its own is not especially insightful, and plenty of people have said something similar. What I think has gone underappreciated is how AI reaches most businesses. Professional services are changing in ways that are hard to see from the outside, which I will come to below. Beyond that, even the physical businesses are exposed, and not only to robots. The AI boom is doing something to the wider economy that changes the cost of capital for every other business on a go-forward basis, and that is the part I think investors are still missing.

It is all AI now. There is no such thing as no AI.

The List Keeps Shrinking

Software went first. The market has spent much of this year working out which products become more valuable as models improve, which turn into features inside somebody else's agent and which were charging a subscription for something that can now be recreated over a weekend.

Professional services looked more reassuring. Lawyers and accountants sell judgement, licensed people still need to sign their names, and clients generally prefer a human explaining their accounts to a regulator.

That logic helped spur a wave of private equity investment into legal and accounting roll-ups. The work looked recurring and fragmented, with plenty of small firms to consolidate.

The defence is weakening. The FT reported in June that private equity executives were warning about the firms they had spent the past few years buying, with Apax identifying hourly bookkeeping as particularly exposed.

The FT's September 1 report on Wall Street pushing Big Law to cut fees found Morgan Stanley and Citi asking major law firms for fee arrangements that pass some of the gains from AI back to clients. Goldman Sachs has been asking firms how much faster the work has become and expects to share in the savings.

Then you reach physical services. A model cannot repair a roof, care for an elderly person or move a pallet through a warehouse.

Whilst that is certainly true today, I think the rate of progress there is going to surprise people. You only need to look at what came out of China's robot games last week, or at Musk's comments on Tesla's Optimus, to see where this is heading. Getting AI to work with robots in real-life situations in an impactful way I’m certain is less than five years away, probably closer to two.

A services business does not need a perfect humanoid competitor to feel the effect. One reliable task in a warehouse or care facility is enough to shorten the list again.

And even if the robots take longer than expected, the capital impact has certainly arrived already.

AI Has Become A Borrower

Microsoft, Google, Amazon and Meta funded the first phase of the boom largely from cash flow. The numbers have grown large enough that the next phase is moving into public bonds, private credit and off-balance-sheet financing.

The Dallas Fed estimates that data-centre investment over the next three to five years may require $3 trillion to $5 trillion. Nvidia has signed agreements with Apollo, Blackstone, BlackRock, Brookfield, Goldman Sachs and KKR to create compute-financing platforms intended to deploy more than $500 billion.

Those institutions finance everything else too.

Brett Winton put the credit-market problem plainly. Compute is scarce enough that data-centre projects can tolerate a higher cost of financing, while the durability of Nvidia chip values allows even balance-sheet-tenuous operators to borrow against the equipment.

High-yield debt will probably feel that competition first because the money already moves towards the best marginal return. The more surprising stage comes when compute is sold to lenders as a stable, investment-grade-like asset and begins competing with companies that are used to rolling their debt at ordinary spreads.

Annual data-centre capital requirements moving into the trillions would be meaningful against total global corporate issuance. A mature business can be performing perfectly well and still discover that lenders have found somewhere more attractive for the next dollar.

The Price Of The Next Dollar

Goldman Sachs estimates that AI investment will reach about $600 billion in the US this year. Its economists think the related debt supply has already added about five basis points to corporate borrowing costs and displaced roughly $10 billion of non-AI investment.

Five basis points sounds small; however, across a large refinancing it can have a meaningful impact on the interest bill, the price a buyer can pay and whether an expansion clears the investment committee.

If stronger AI-driven growth pushes long-term rates higher, the pressure compounds. Technological deflation may keep parts of inflation in check, but a stable company that relies on cheap refinancing can still find its operating model under pressure when the base rate and its spread rise together.

Whilst crowding out might not be a realistic concern right now, I think the direction of travel is meaningful. The effect has already moved beyond technology and into the borrowing cost of the rest of corporate America.

A Recurring Claim On Capital

PwC's global outlook projects $31.6 trillion of data-centre capital expenditure through 2050. Long forecasts deserve scepticism, but the shape of the spending matters more than the final number, because this is not a one-and-done buildout.

The great infrastructure programmes of the past needed one enormous upfront investment and then ran for decades. A railway or a power station has a 20 to 30 year life. A data centre is different: most of its cost sits in servers, networking equipment and GPUs that get replaced every four years or so, and PwC estimates every dollar of construction commits the market to about twelve dollars of future technology spending.

So the claim on capital never really ends. If demand holds, the same sites come back to the market for new equipment every few years. If it fails, lenders are left with collateral that depreciated faster than the debt. Either way, the pools of capital that finance the rest of the economy are being asked for money on a cycle far shorter than anything the infrastructure playbook was built for.

The Physical Hurdle Rate

Data centres also need power, grid connections, cooling equipment, fibre, land and construction workers. A manufacturer may never compete with ChatGPT for customers, but it can compete with a data centre for a transformer, a contractor or the next gas turbine.

Goldman's economists found gross margins on data-centre construction are more than twice those on non-technology projects, so the builders and their crews go where the money is.

Planning adds another constraint. Gillian Tett's FT column on the data-centre backlash reported that just 20 per cent of Americans in a YouGov poll said they liked data centres. Local opposition delays sites and concentrates construction in the places still willing to approve it.

The political response will determine where capacity gets built, but it will not remove the underlying demand. Scarce sites become more valuable, power queues lengthen and nearby industrial projects wait.

AI Risk Is Becoming Everything Risk

Deal teams have tended to assess AI risk through the income statement: jobs that can be automated, products that can be replicated and customers that might leave. The data-centre boom puts the balance sheet and capex plan into the same discussion.

For years, investors treated technology exposure like China risk or Amazon risk, a discrete item to score and sometimes ignore. AI exposure is becoming universal.

Calling it a risk is almost misleading because risk implies the event might not happen. The buildout is under way; what varies is whether it reaches a business through its customers, workers, borrowing costs or the physical inputs it needs to grow.

So when someone asks which industries are insulated, I still struggle to name one.

What The Builders Are Saying

Two posts worth reading this week.

@Alfouille (AI automation builder and cybersecurity trainer)

The post: staff were using personal AI accounts, sometimes with company data, and nobody knew what happened to the information afterwards. The AI meeting ended with a rather large clean-up job.

Why this matters: the rollout happened before the policy. A buyer inherits the data leakage and the hidden dependence on tools the target may not know it uses.

My take: AI policy now belongs in diligence alongside the security controls. I would compare the approved-tool list with expense records, browser extensions and SSO logs, because the gap between policy and actual usage is where the interesting risk sits.

@ttunguz (Tomasz Tunguz, investor)

The post: Tunguz estimates that a roughly $5 trillion AI buildout could create $4 trillion of debt and $260 billion to $300 billion of annual interest. His servicing maths requires $1.2 trillion to $1.5 trillion of annual AI revenue by 2030.

Why this matters: if the revenue misses, lenders own GPUs that age on a software timetable. If it arrives, trillions of dollars still move away from other borrowers.

My take: I would add one column to every refinancing schedule: the spread at which the investment case stops working. A business does not need AI exposure in its revenue line to become vulnerable when its debt rolls.

News Digest

Nvidia Buys The Open-Weight Distribution Layer

Nvidia has agreed to buy Hugging Face for roughly $13 billion. Hugging Face began as an emoji-named chatbot app and became the main repository, community and distribution point for open AI models. It now sits inside the company selling most of the chips on which those models run, and every extra model ecosystem creates another reason to buy compute.

The details:

Why it matters: Nvidia is buying the marketplace where developers choose models, giving it influence over demand before the order reaches a chip distributor.

My take: the deal works while Hugging Face still feels neutral. Nvidia makes money whenever more models need compute, but developers can move if the repository steers them too aggressively. I would underwrite the community separately from the code.

Read more

Morgan Stanley And Goldman Get The Anthropic Mandate

The FT reported on September 7 that Anthropic is close to appointing Morgan Stanley as lead-left bank and Goldman Sachs as stabilisation agent on an IPO its shareholders expect to price at $2 trillion or more. JPMorgan, Citi and Barclays are in line for senior roles after providing debt financing. The prospectus is now expected in late September, with marketing beginning in mid-October at the earliest, according to Reuters.

The details:

Why it matters: the largest IPO in history arrives with a compute bill that is itself one of the debt stories in the FTT.

My take: I am interested to see what the S-1 says beyond what we already know about the impact on knowledge work and coding. The impact on the physical world is still massively underpriced, and arguably represents an even bigger opportunity.

Read more

Other Interesting Things I've Read or Seen This Week

Private credit marks down its software loans again (Reuters, September 3) - Across 44 BDCs, 81 per cent of software loans have been written down this year against 40 per cent elsewhere, and non-accruals rose to 3.4 per cent of cost. (The SaaS panic has left the equity market and moved into the loan book.)

Jefferies sets out the new underwriting standard for software private equity (Jefferies, September 1) - Buyers now want evidence of AI defensibility and structures with downside protection, from earnouts to liquidation preferences. (The term sheet has become the place to express doubt.)

Big Tech profits get a $160 billion boost from stakes in other AI companies (FT, September 2) - Alphabet booked $97.9 billion of other income last quarter and Amazon $53.4 billion, largely paper gains on AI holdings and the SpaceX listing. (Circular financing now has its own line in the income statement.)

Charter's CFO leaves for the Blackstone-Google data-centre venture (Fortune, September 3) - Jessica Fischer becomes founding CFO of a venture backed by $5 billion of Blackstone equity and targeting 500MW by 2027. (The FTT, in a single career move.)

Microsoft reorganises its reporting around Agents and Infra (WSJ, September 2) - Three reporting segments become two: Agents and Infra, plus Devices and Consumer. (When the org chart reaches the 10-K, the strategy has become rather official.)

The $53 billion PayPal takeover fell apart over who owned the share-price rise (Axios, September 1) - Stripe and Advent offered $60.50 a share, PayPal traded through it, and both sides claimed credit. (Price discovery does occasionally work.)

McKinsey finds the gap between individual gains and enterprise impact (McKinsey, September 1) - 80 per cent of respondents say AI improved their own productivity, while the share reporting an EBIT contribution is stuck at 37 per cent, unchanged on the year. (Last week's FTT, with a sample size.)

Wonderful raises $550 million at a $5 billion valuation (WSJ, September 2) - The year-old enterprise-AI startup already employs 650 people and plans to build a 1,000-strong team. (Software may be eating the world, but apparently it still needs quite a large implementation department.)

OpenAI releases Astra and admits it sometimes tries to evade monitoring (Reuters, September 3) - OpenAI says its new model is faster and more capable, with strengthened safeguards following the Hugging Face incident. (The release notes have become rather more consequential.)

Tim Cook hands Apple to John Ternus (Reuters, September 1) - Ternus inherits a $4.6 trillion company whose position in AI is considerably less settled than its position in phones. (A fairly comfortable first day, apart from the difficult bit.)

Acquisition Intelligence is a weekly newsletter on AI in M&A for finance professionals, private equity investors, investment bankers, corp dev teams, and deal-makers.

For questions, feedback, or to share what you're seeing in the market, reply to this email.

P.S. I'm Harry, co-founder of DealSage. We help investment firms work out how AI changes the deal in front of them, including risks missed by a generic automation checklist. Reply here if you would like to look at one of yours, or have a look at dealsage.ai.