IBM, and the Plumbing Enterprise AI Runs On

This is a separate note from the short-dated IBM options trade we published in September and closed the same month. That was a two-week bet on drift. This is the opposite: a long position in IBM common stock, bought on 23 October 2026 for a three-to-five-year hold, on a thesis that has nothing to do with the tape and everything to do with where enterprise AI actually gets stuck. We are keeping the old trade note published, closed and labelled, because we do not delete the record. Treat this as the real coverage.

The thesis in one line: the bottleneck in enterprise AI is not the model, it is everything around the model, and IBM owns an unusual amount of that everything. Getting a large company from an AI demo to AI in production means integrating it into systems of record that are decades old, governing it so a regulator and a board will sign off, running it across a hybrid of on-premises and multiple clouds, and redesigning the actual workflow so a human does less and the machine does more. None of that is glamorous. All of it is hard, slow and sticky, and it is exactly the layer IBM has spent the Red Hat era rebuilding itself around: OpenShift as the hybrid substrate, watsonx for governance, Consulting to do the workflow redesign, and the mainframe and transaction systems that still hold the data the AI needs to touch.

We are buying it now because the market just handed us the entry. IBM reported a genuinely good quarter on 21 October, then fell about 7% the next morning because the CEO said enterprises are getting disciplined about the return on their AI spending. We think the market sold the wrong sentence. ROI discipline is not a headwind for the company that sells integration, governance and workflow redesign; it is the tailwind, because it is what happens right before a buyer stops running a hundred experiments and starts paying someone to put two of them into production. By the close on 23 October the stock had already recovered to $253.56, which is where we opened. We are not calling the bottom of a two-day swing. We are saying the business under it compounds, and the panic told us what the marginal seller is afraid of, which is the thing that helps IBM.

IBM is no longer a hardware company with a software arm. In the third quarter of 2026 Software was $7.8 billion of the $17.1 billion total and grew 9%, and its segment profit grew 15% at a 35% margin. That is the engine. Underneath it, the recurring base is what matters for a multi-year holder: software annual recurring revenue reached $24.9 billion, up 8%, and within it the pieces tied to the thesis are the fastest. Red Hat grew 11% and OpenShift's ARR grew more than 30%. Data grew 20%. Transaction Processing, the software that runs on the mainframe and that everyone has written off for twenty years, grew 10%.

Consulting is $5.3 billion and grew 2%, and it is the soft spot we discuss honestly in the risk section. Infrastructure is $3.7 billion, up 5% at constant currency, carried by storage and distributed systems even as the IBM Z mainframe line fell 19% as the z17 cycle matures, having added more than a billion dollars of incremental revenue over the prior generation on the way up.

The shape to hold in your head is a company that is roughly 45% software, high-margin and recurring, bolted onto a consulting arm that does the implementation and an infrastructure arm that owns the systems of record. That combination is not something a hyperscaler or a model lab can easily assemble, and it is precisely what an enterprise needs when it stops buying AI as a toy and starts buying it as plumbing.

Here is the house question applied to IBM: where is the structural constraint in enterprise AI, and who owns it? The constraint is not access to models, which are now a commodity sold by the token. The constraint is everything between a working model and a deployed workflow inside a real company, and it has four parts, each of which is a line of IBM's business.

Integration. The valuable data lives in systems of record that are old, regulated and not going anywhere. A bank's core runs on a mainframe. Getting an AI to read and write against that safely is integration work, and IBM's Transaction Processing software growing 10% in 2026 is the tell that the mainframe is not the past but the place the data sits.

Hybrid deployment. Most large enterprises will not put their crown-jewel data in one public cloud, so AI has to run across on-premises and several clouds at once. That is what Red Hat OpenShift is, and OpenShift ARR growing more than 30% is the clearest single number in the quarter. It is the substrate the AI runs on, and IBM owns it.

Governance. A regulated enterprise cannot deploy a model it cannot audit, explain or control. That is watsonx's governance pitch, and it becomes mandatory, not optional, exactly when boards start asking about the "true ROI" the IBM CEO flagged.

Workflow redesign. The ROI only shows up when the work itself changes, and redesigning a claims process or a supply chain around an agent is consulting work. Generative AI is now embedded in engagements representing 35% of IBM Consulting's backlog, which is how the AI story shows up in a services business: not as a product line, but as the thing the backlog is increasingly about.

No single competitor owns all four. The model labs own the model and almost none of the surrounding work. The hyperscalers own one cloud each and want the data in it. The pure consultancies own the people but not the software substrate. IBM is the one large company that sells all four layers of the plumbing, and the plumbing is the bottleneck.

On the earnings call the CEO said enterprise customers are "a lot more cognizant and a lot more aware of what is the true ROI of all of their AI spend," and that some are nervous about "AI spending completely out of control with no real return." The stock fell about 7% the next morning, and the read-across hit the consulting names too. We think that reaction inverts the logic.

A buyer who is scrutinising ROI is a buyer who is about to consolidate. The hundred small experiments, run on a hundred model endpoints with no governance and no workflow change, are the thing that gets cut. What survives is the two or three use cases someone is willing to put into production, and putting something into production is the integration, hybrid, governance and consulting work IBM sells. The CEO said as much and the market did not listen: he called the demand for workflow redesign and token-cost control "the tailwind for us," and urged firms to "pick a few" initiatives rather than run a hundred. "Pick a few and put them into production" is the most bullish possible sentence for IBM's actual business, delivered on the day the stock fell for it.

We are not claiming the quarter was flawless or that the recovery to $253.56 by the close proves anything in two days. We are claiming that the fear the selloff expressed, that the AI experimentation boom is cooling, describes a rotation from experimentation toward production, and production is where IBM gets paid. That is the kind of mispricing a multi-year holder is supposed to buy.

This is a cash machine, which is what makes a three-to-five-year hold tolerable while the thesis plays out. IBM generated $3.03 billion of free cash flow in the quarter and raised its full-year free cash flow outlook, guiding to an increase of more than a billion dollars. Pre-tax operating margin was 19.3%, up 70 basis points, so the software mix shift is showing up where it should, in margins. Adjusted earnings were $2.95 against a $2.87 estimate.

The balance sheet supports a dividend of $6.76 a share, a yield of about 2.7% at our entry, which IBM has raised for three decades and which free cash flow covers comfortably. For a holder, the math is simple and unglamorous: a mid-single-digit revenue grower with an expanding margin, a growing free cash flow line, and a near-3% dividend, where the recurring software base ($24.9 billion of ARR) is both the largest and the fastest-growing part. You are paid to wait, and the waiting is funded.

The one thing management did not do was raise the full-year revenue growth guide, which stayed at 4% to 5% constant currency, and that restraint is part of why the stock fell. We read a company that beats on margin and cash while keeping the top-line guide conservative as the better of the two failure modes.

IBM is not a bargain and we will not pretend it is. At $253.56 it trades at about 21 times trailing and roughly 19 times forward earnings, with a market capitalisation near $239 billion. That is a market multiple for a company growing revenue in the mid-single digits. The bull case is not that the multiple is cheap; it is that the mix behind the multiple is improving, and the market is still paying a blended hardware-and-consulting multiple for a business that is becoming a software-and-recurring one.

The way this compounds is not multiple expansion we are counting on, but the combination underneath: software ARR growing 8% and mixing up, margins expanding as software becomes a larger share, free cash flow growing, and a 2.7% dividend paid while it happens. If software keeps mixing up and the market eventually pays a software multiple for the software half, that is the upside, and it is optional rather than required. Required is only that the recurring base keeps compounding and the cash keeps coming, which it is.

Buy IBM expecting a quick re-rate and you have the wrong note. Buy it as a cash-generative compounder quietly changing its own mix, bought on a day the market mistook ROI discipline for a problem, and the entry is reasonable.

We are long IBM common stock, opened on 23 October 2026, for a three-to-five-year horizon. We hold shares, not options; this is the structural position the September trade note explicitly was not. The disclosure block at the top of this note carries the terms, and we will update it as a dated entry if anything changes.

One honest note on timing: we are buying immediately after an earnings event, into a stock that swung roughly seven percent intraday and then recovered within two sessions. That volatility is noise against a multi-year thesis, and we would rather enter while the marginal seller is still reacting to a sentence we think they misread than wait for the tape to calm down and pay up for the calm. Sizing is Constructive: a real position taken with conviction, but one we expect to add to on weakness given that the entry multiple already reflects a decent business.

The thesis has three specific risks and one general one, and the first is the mirror of the bull case.

Consulting is the soft spot, and it is where the AI knife cuts both ways. Consulting grew only 2%, signings fell 2%, and segment margin slipped 60 basis points. The same enterprise ROI discipline we are calling a tailwind is, in the near term, a reason buyers delay large consulting engagements, and there is a real bear case that generative AI eventually compresses the billable-hours model that consulting is built on. We are underwriting that IBM's consulting becomes the delivery arm for AI deployment faster than AI erodes its traditional work. That is a genuine open question, not a settled one.

The mainframe cycle is rolling over. IBM Z fell 19% as the z17 cycle matures. The installed base is sticky and Transaction Processing software keeps growing, but a holder should expect the hardware line to be a drag until the next cycle, and should not mistake the software that runs on the mainframe for the mainframe itself.

IBM has been a value trap before. For most of the 2010s IBM told a transformation story while revenue shrank and the stock went nowhere. The counter is that the software mix and recurring base are now genuinely different, and that free cash flow is growing rather than being financial-engineered, but anyone buying IBM for the long term is buying against a decade of disappointed patience, and should size accordingly.

The general risk is valuation: at 19 times forward earnings there is little margin for a growth stumble, and a broad rotation out of large-cap tech, or a couple of soft consulting quarters, could leave the stock lower a year from now while the three-to-five-year thesis stays intact. That is the scenario we are most prepared to sit through.

The open questions we consider the next highest-value work on this name.

1. 01

   Track OpenShift and software ARR, not revenue. The thesis lives in the recurring base and the hybrid substrate. OpenShift ARR growth and total software ARR are the numbers that confirm or break it, quarter by quarter.

2. 02

   Watch consulting signings and book-to-bill. A trailing book-to-bill of 1.04 is barely above one. If AI deployment work is real, signings should re-accelerate; if AI is eating consulting, they will not. This is the single most important swing factor.

3. 03

   Quantify the gen-AI backlog. "35% of consulting backlog" is a share, not a dollar figure, and IBM declines to give a standalone number. Build an estimate and track whether that share is rising.

4. 04

   Model the free-cash-flow bridge. The dividend and the thesis both rest on FCF growth. Separate genuine operating cash generation from working-capital and one-off effects.

5. 05

   Test the consulting-disruption bear case directly. Model what IBM looks like if AI compresses consulting margins rather than expanding the deployment opportunity. That is the scenario that would break this note.

• IBM Q3 2026 results, reported 21 October 2026
• IBM Q3 2026 earnings call transcript
• Software, Red Hat and OpenShift segment detail
• Consulting signings, backlog and book-to-bill
• IBM Z and Transaction Processing disclosures
• CEO remarks on enterprise AI ROI
• IBM free cash flow guidance
• IBM price, market capitalization and dividend
• IBM short interest and ownership
• watsonx governance and hybrid cloud
• Analyst reactions to the quarter