When AI makes software cheaper to build and business processes easier to automate, who gets to keep the profits?

For most of the last thirty years, software has been an extraordinary business.

Build a product once, Sell it to thousands of customers, Charge recurring fees, Keep the incremental cost of serving another customer relatively low. That’s pretty much the recipe! SaaS of course made this model even more attractive.

AI is challenging the economics behind this model in three ways.

1. It is making software cheaper to produce.

2. It is allowing customers to use applications without interacting with their interfaces.

3. It is allowing software to perform work rather than merely help people perform it.

While I do expect the total software market to grow significantly, I don’t expect the existing software companies to capture that growth in proportion to their current market positions.

My bet is that the greatest opportunity for profit will shift from selling software to taking responsibility for getting work done.

Software is becoming cheaper to reproduce

AI doesn’t make enterprise software free. Security, reliability, scalability and maintenance remain difficult. But functionality that once required a large engineering organization can increasingly be developed by much smaller teams.

A software feature is valuable partly because of what it does and partly because of how difficult it is for someone else to reproduce. When reproduction becomes cheaper, the second source of value weakens.

What happens when customers stop using the application?

Suppose I want to approve an employee’s promotion. Today I might log into an HR application, navigate through several screens, review compensation data and submit an approval. Tomorrow I might ask an enterprise agent to review the case, compare compensation, verify the budget and prepare a recommendation. Once I approve it, the agent completes the transaction.

The HR application remains important. It holds employee records, organizational structures and compensation history. But I may never open it.

This is the logic behind the argument that SaaS companies should expose their functionality through standards such as MCP and allow ChatGPT, Claude or Copilot to handle the interaction. I agree with the architectural direction – but I am not convinced by the economics.

Say a vendor earns $100 per employee per month. If customers increasingly consume those capabilities through external agents, why should the vendor automatically retain the same price? Gross margins could improve. But higher gross margins on substantially lower revenue aren’t necessarily a better business.

Of course, established vendors won’t just passively surrender all of that revenue. They can charge for API calls, introduce agent licenses, bundle their own agents with the application or restrict access to certain capabilities. Those are credible defenses, especially when switching costs are high.

A standard like MCP isn’t a real business moat, and neither is owning the agent that calls it.

Why I wouldn’t bet on orchestration alone

A common prediction is that value moves from individual applications to platforms that coordinate them. There is some truth to this. But why should the agent platform retain attractive margins?

Model providers are building agents. Hyperscalers have infrastructure, distribution and enterprise relationships. If multiple platforms can call the same tools and execute the same workflows, the orchestrator can become interchangeable too.

The difficult part is not invoking an application. It is knowing what should happen, under which circumstances, with whose permission, and what to do when something goes wrong.

The system of record is not the system of authority

Picture a bank investigating a suspicious transaction. The core banking application records it. A financial crime system identifies unusual activity. An agent gathers evidence and recommends whether to block it.

Who makes the decision?

Today, the bank’s policies establish the criteria, authorized employees or systems exercise the decision rights, and the bank remains accountable to its regulators.

AI doesn’t eliminate that structure. But it changes how much of the work can be delegated. The bank could authorize an external platform to execute predefined actions within approved policies and escalate decisions requiring human authorization. That platform would hold delegated decision rights, but not unrestricted authority.

The bank would retain ultimate regulatory accountability. The vendor could accept contractual liability for failures within its agreed responsibilities.

Decision rights can be delegated. Accountability cannot always be transferred.

In Migrating Scarcity, I argued that when intelligence becomes abundant, scarcity shifts toward the ability of institutions to absorb and authorize it. That becomes a central economic question for enterprise software.

When software does the work, pricing has to change

Take a company with 1,000 people processing invoices. Its software costs $100 per employee per month, so the vendor earns $1.2 million annually. Now assume automation allows 200 people to handle the original workload. Under unchanged seat-based pricing, the vendor’s revenue falls to $240,000.

The customer is substantially more productive. Yet the vendor has lost 80% of its revenue.

But the vendor doesn’t inevitably lose 80%. It could charge per invoice. And lower processing costs could create demand for more automation. Human users could decline while total software consumption increases.

Still, changing the pricing metric isn’t the same as changing what the company sells. Charging per invoice is still selling software if the customer remains responsible for the operation. Taking responsibility for processing those invoices correctly is a different business.

Isn’t that just outsourcing?

Taking responsibility for outcomes sounds like business process outsourcing, and there is considerable overlap.

The difference lies in how the operating capability scales. Traditional outsourcing often required additional people as transaction volumes increased. An AI-native operations provider has the opportunity to separate transaction growth from headcount growth.

But that opportunity depends on something very specific. How much work can the system complete without human intervention, and how expensive is the work it cannot?

The economics of exceptions

Imagine an illustrative AI-native operation charging $2 per transaction. A routine transaction costs $0.20 to execute. One requiring human investigation costs $12. Other operating costs add $0.40.

At 95% straight-through processing, average cost is about $1.19, a contribution margin of 40.5%. At 90%, cost rises to $1.78 and contribution margin falls to 11%.

Nothing changed about the selling price or the AI model. Only the exception rate changed. The last few percentage points of automation can determine whether the economics work.

My hypothesis is that mature, standardized operations with sufficiently high straight-through processing rates can achieve 15–30% operating margins. Contribution margins exclude product development, sales and corporate overhead, so these are different measures. The attraction isn’t SaaS-style margins. It’s that a provider might earn attractive operating margins while capturing a much larger portion of the customer’s total spending on the process.

The profit opportunity comes from spreading the cost of building and controlling the operation across many customers, while keeping the incremental cost of execution and exceptions low.

Why this won’t simply commoditize

This is the obvious challenge to my argument. If model providers, hyperscalers and SaaS incumbents can all offer autonomous operations, why should any operating platform retain pricing power?

Some won’t. Routine, standardized work with low switching costs will face price competition, and providers that simply package an agent with a few APIs will struggle to defend margins.

The stronger positions are in processes where reliability is difficult to establish and failures are consequential. A provider that has handled millions of cases under agreed controls has evidence of competence, and replacing it means retesting controls and changing contractual responsibilities.

But reliability alone isn’t a permanent moat. The advantage comes from continuously improving the economics of execution while maintaining that reliability. A competitor may reproduce the technology. Matching the operating performance across thousands of exceptions, changing customer requirements and regulatory obligations is a different challenge.

What would prove me wrong?

I would reconsider in three cases. First, if conventional SaaS vendors retain their pricing power despite declining human usage, without assuming greater responsibility for business outcomes. Second, if general-purpose orchestrators capture most incremental enterprise automation profits. Third, if productized operating businesses fail to generate attractive returns after accounting for exceptions and liability.

If I ran a software company, I would ask three questions.

1. Which parts of our revenue depend on users, which on transactions, and which on capabilities customers cannot easily replace?
2. What decision rights would customers trust us to exercise, and what liability are we prepared to accept?
3. Does our business become more valuable when our customers become more productive?

If the answer to the last is no, adding AI features won’t fix it.

For thirty years, software companies built great businesses by selling tools that people used to do work. The next generation will increasingly be built around taking responsibility for getting the work done.

---

Nvidia announced Jensen Huang’s resignation on Thursday, naming CFO Colette Kress interim CEO and Harvey Jones its first-ever board chair (Nvidia 8-K). It followed an independent board investigation announced on November 8 (AI Business). Of all the questions this raises, the one that interests me most is what boards can learn from it.

Imagine being a director at Nvidia and telling Jensen he needs to change the way he runs the company. Here is a man who created several trillion dollars of shareholder value doing things conventional management wisdom would have advised against. And yet, that is part of the job shareholders pay directors to do. The board needs a point of view on how the company operates, not just how it performs.

Take the span of control. Huang reportedly had as many as 60 direct reports (Yahoo Finance). I have no issue with that if he could make it work, and clearly he did. But how many of those executives got to make the kinds of decisions Jensen made?

You can delegate a lot of work without delegating much authority. And without exercising authority, it is hard to develop the judgment needed to be a CEO. I explored this in The Corner Case. A flat structure can be fast and still fail to produce a successor.

So what should boards do?

Discuss succession at least twice a year, with and without the CEO present. There should be a plan for tomorrow morning and another for three to five years from now.

Know the bench firsthand. Directors should spend time with the next two levels of management, independently assess potential successors, and influence the assignments they get. They should also know which customer, partner and government relationships depend disproportionately on the incumbent CEO.

Tell shareholders the process. The board doesn’t need to name potential successors, but it should disclose how succession is handled. Semafor reported that the word “succession” appears zero times in Nvidia’s latest annual report, against 25 in Microsoft’s (Semafor via AI Weekly). Huang himself said the next CEO has to be hired, just as the board could fire him. That is the right principle. It just needs a visible process behind it.

Plan for the board itself. Why stop at the CEO? How many boards have a good succession plan for themselves? Huang, Tench Coxe and Harvey Jones have all sat on Nvidia’s board since 1993 (The Next Web), and average tenure is nearly 15 years, roughly double the S&P 500 norm. Long tenure brings institutional memory that is hard to replace. But refreshment isn’t only about adding skills. It is also about keeping the ability to challenge a founder. Directors who have spent decades alongside a brilliant CEO find that harder to do. Bloomberg noted that Nvidia’s board has no current or former S&P 500 CEOs (Bloomberg). Tenure norms and a skills matrix reviewed every year would help.

Build enough AI fluency to challenge management. Many boards are being asked to approve billions in investment and fundamental changes to business models in an area where few directors have firsthand experience. I don’t expect every director to become an AI expert. I do expect the board collectively to have enough expertise to challenge management without relying on management to educate it. A director should be able to ask which parts of the company’s competitive advantage will survive when its competitors have access to the same AI capabilities, and then independently assess the answer. That question sits at the heart of Migrating Scarcity.

Watch where AI moves decision-making. As AI makes execution easier to scale, companies may concentrate authority in fewer hands. But AI can just as easily spread decision-making to more people, because it puts analysis and judgment tools in front of people who never had them. The economics of either path can be attractive. The governance implications differ. The board should understand which direction management is choosing, and why.

To its credit, Nvidia named an experienced internal executive as interim CEO very quickly. We don’t know enough about its long-term succession planning to judge the board conclusively.

But I would encourage every board to ask itself two questions at its next meeting.

1. If our CEO left tomorrow, how much of our competitive advantage would leave with them?

2. And if the answer makes us uncomfortable, why are we waiting?

---

America has a commanding lead in AI compute. Whether that becomes a lasting economic advantage is a separate question.

There is a growing argument that America loses the AI race to China if we don’t build data centers faster. I think there is a real concern behind it. Compute is a constraint, power takes years to build, and China is investing heavily in energy and industry.

But building more data centers is not a strategy by itself. So, are we measuring the wrong things? Partly, yes. We keep count of capacity: chips, megawatts, data centers. We count results far less.

I find it easier to think about three separate questions.

1. How much capacity does America need?

2. Who pays for it?

3. And how well can our economy turn it into productive activity?

The third one gets the least attention in the public debate.

America is ahead. By a lot.

The Lowy Institute’s November analysis concludes that America is winning. It estimates that US AI data centers have roughly 40 times China’s computing power and US AI chips 31 times, although the report doesn’t spell out how those multiples were built. American models draw more than 90% of combined web traffic to US and Chinese AI services. Lowy also says its model rankings are consistent with claims that the US is six to nine months ahead, a figure it cites but does not measure itself.

That strengthens the case for keeping our compute capacity growing.

Now put two numbers from the 2026 Stanford AI Index next to each other. US private AI investment in 2025 was $285.9 billion, against $12.4 billion in China. And in March 2026 the best American model led the best Chinese model by 2.7% on the benchmark Stanford highlights.

These aren’t like-for-like. Chinese state funding isn’t in the private investment figures, a benchmark covers a narrow slice of what models do, and capital spending pays back over years.

Still, it makes me ask how much capability and value we get from each additional dollar invested.

China’s progress doesn’t show compute is unimportant. Chip export controls have made it a real constraint for Chinese developers. But Chinese labs have shown that efficiency can close part of the gap. We should care about how well we use compute, not only how much we have.

Open models

China is strong in open-weight models. Per Lowy, Chinese models were downloaded about 1.8 billion times on Hugging Face in the third quarter of 2026, against 500 million for US models. That is 78% of the two countries’ combined total.

Downloads aren’t deployments, though. Many are small models and automated pulls, and American models still dominate consumer use.

Lowy also reports that Chinese models handled 55% of text requests on OpenRouter during one week in late October, against 38% for American models. That is one developer marketplace for one week, and Lowy itself notes that most people reach AI through apps and websites, where US models dominate. It is one signal of where developers are choosing to build.

A company that can run an open model on its own hardware and tune it may get good results at much lower cost than a large proprietary model would give. That is worth supporting, along with independent testing of provenance, security and behavior.

Neither open weights nor a model’s country of origin tells you whether it can be trusted.

The real constraint is speed to power

Most large training clusters don’t need to sit near customers. Some inference does. Developers already look at remote sites and their own generation.

The bottleneck is how fast they can get power, and new transmission in advanced economies can take four to eight years.

FERC moved in June, ordering six regional grid operators to justify or reform their rules for connecting large loads. The first responses are due November 16. Those will be proposals. The real test is whether connection times fall once new rules are in force.

In the meantime, dedicated generation may be the fastest option for some projects, even if it is more expensive or less environmentally attractive. Modernizing the grid is necessary, but it won’t solve the next two years of demand.

I’d like to see published connection timelines, utilities saying what each new large load requires, and a transparent split of incremental costs. Grid upgrades help manufacturing too, so we shouldn’t treat them as a subsidy for AI.

Water gets a lot of attention. Closed-loop cooling cuts water use but can raise electricity use, so the trade-off has to be checked site by site. In some places the power bill will matter more. In others, water availability will be the bigger concern.

Who pays if the forecasts are wrong?

Companies are committing hundreds of billions of dollars based on forecasts of demand, prices and hardware performance. Some of those forecasts will be wrong.

Cheaper inference could shrink the compute each task needs, or it could grow total demand. Both are plausible.

If a utility builds for a data center that gets cancelled, or that uses half the power it forecast, existing customers shouldn’t pick up the bill.

Ohio’s large-load tariff uses minimum payments. State commissions should look at long-term commitments, financial guarantees and exit charges, without piling on so much that good projects die.

Investors should keep the upside and carry a fair share of the downside.

The constraint nobody is measuring

Gartner expects cloud spending on running AI models ($23.3 billion) to pass spending on training them ($19 billion) this year.

The 2026 McKinsey survey says 80% of respondents report personal productivity gains, but only 37% report any positive effect on enterprise EBIT, and about 6% qualify as high performers, attributing at least 5% of EBIT to AI with significant overall impact.

The last two numbers are flat from 2025, even though adoption and spending went up.

Spending on running models is climbing, yet the share of surveyed companies reporting an enterprise profit effect isn’t. That doesn’t prove infrastructure is being overbuilt. Returns may be lagging the investment, or may be concentrated in a small group of companies. Either way, making intelligence more available isn’t automatically producing broader financial returns.

McKinsey also found that nearly three-quarters of high performers rebuilt workflows around AI, against about a quarter of everyone else. The survey can’t show cause, but it matches kind of what I’ve seen in enterprise work.

A bank automates document review and leaves credit approval alone. An insurer speeds up claims and keeps decision rights where they were. The AI works. The organization hasn’t changed enough to capture the value.

That is the argument in Migrating Scarcity. When intelligence gets abundant, scarcity moves to the next constraint, which is our ability to absorb it, authorize its use and redesign work around it.

Is China better at this? I don’t have the evidence to say. Stanford reports that China installed about 295,000 industrial robots in 2024, 54% of the world total, against about 34,000 in the US. That mostly reflects the size of its manufacturing base, but it is a reminder that deployment matters as much as model quality.

What I’d check in 2028

Enterprise returns. Move the share reporting any EBIT gain from current baseline of 37% that McKinsey reports to say 50%, and double the high performers from 6% to say 12%. These are targets I’d use to judge progress, not predictions. McKinsey’s sample is global, so I’d also want a US-only cut.

Grid timing. A one-third cut in the median time from a complete application to energization, measured against a published baseline. Alongside it, publish cancellations and any costs shifted onto existing customers.

Cost per task. Not the price per token. The cost of finishing a defined set of tasks at fixed quality, reliability and security, at least halved from today. This needs to be baselined quickly – I am surprised this hasn’t been done yet.

New things. New products, markets and companies that couldn’t have existed before AI. This is harder to count, but I’d start with whether business formation in AI-related sectors is growing faster than business formation overall.

What would worry me

If we miss the enterprise-returns targets while capacity keeps expanding, our ability to use AI isn’t keeping up with our ability to produce it. That wouldn’t necessarily mean we built too much. But it would make me question where the next dollar of investment should go.

I’d also want to know whether data center owners are earning adequate returns. Enterprise productivity alone doesn’t guarantee they made good investments.

America should build the compute it needs, modernize the grid and protect ratepayers. None of that removes the need to change the institutions that will use the intelligence.

As I argued in AI and the Next American Economy, we also need new products and industries, not only cheaper versions of today’s work.

Compute is the lead we already have. Using it productively is how we keep it.