Allow me to present a controversial (but logical) premise. That means the prices we pay for AI tools will start to rise. And this price increase will have some very positive (and some distracting) effects.
It’s very simple. This is a very expensive technology to provide. Already, data center spending, adjusted for inflation, far exceeds the cost of building an entire 47,000-mile highway system over 40 years ($670 billion).
Here are my investments over the past 12 months, which I think are significantly undervalued.
big 4 hyperscalers — Amazon, Alphabet, Microsoft, Meta: approximately $370 billion to $410 billion in 2025, depending on whether strict capex, finance leases, or fiscal year adjustments are used. Reuters cited Bridgewater estimates that these four companies are expected to invest about $410 billion in 2025 and about $650 billion in 2026.
Add in Oracle, CoreWeave, and xAI/SpaceX AI infrastructure, and the world of working “AI data center builders” is now around $500 billion. recent Investments are annualized and trend toward more than $700 billion to $750 billion in run-rate spending in 2026. The broader market, including Stargate-style multi-year contracts, is much larger, but these numbers should be treated as committed or announced capacity rather than used capital. x
Add in companies like Nvidia, TSMC, Micron, Intel, SK Hynix, and Seagate and you could easily get another $200-300 billion, pushing the run rate in 2026 to nearly $1 trillion.
And it gets even worse. Gartner predicts this will be $6.3 trillion by 2030.
 |
As many new companies go public (Anthropic, OpenAI), they will be under pressure to show positive gross margins (Anthropic is close), which will drive prices up. And all the SaaSapocalypse companies (SAP, Workday, Oracle, Salesforce, Adobe) will also want to show Wall Street that they are making money.
This week I was in New York with clients and heard three times that CIOs and CHROs were already considering whether they should “outsource” their AI to engineers in India because of the high cost of code.
takeout (From information)
Eric Johnson, chief information officer at PagerDuty, which helps software engineers navigate technology failures, said the company’s 1,200 employees have begun using Anthropic’s AI coding and other tools to speed up software development and other tasks, bracing for volatile costs.
 |
“I’m prepared to be surprised” by the bill, the CIO said. “We believe there’s a lot of value here. Unfortunately, this is a fairly new technology, so there are some open questions about cost and return on investment that we’re going to answer.”
Companies with employees who are heavy users of Anthropic’s Claude product could end up paying significantly more for the product after the company changed its pricing model to charge enterprise customers based on the amount of AI they use, rather than charging a flat fee. Anthropic said it uses a new version of a technology called a tokenizer in its latest AI models, which could also contribute to higher costs for customers.
Many technology companies and Anthropic’s large customers say they are trying to eat into rising costs by making software engineers and salespeople more productive by automating certain tasks.
Update: The Gemini 3.5 Flash, just announced this week, is probably 10 times less expensive than the Opus 4.7, so the battle for price performance has officially begun.
Second, how much will prices go up?
Please read it carefully.
The total “new revenue” that would need to be earned to generate a compounded return of 15% (assuming a generous 5-year depreciation expense) is: $1 trillion annually. Based on AI margins, it’s probably more.
I think some of this revenue will come from consumers and advertising, and some from businesses.
On the consumer side, Internet ad spending today is around $750 billion or so, which includes all advertising on all platforms. So if we More than double the amount of junk ads Well, these companies could almost pay this off.
On the business side, all enterprise software spending is around $1.2 trillion (Gartner); You can also double it.
No matter how you look at it, someone (namely us) will end up paying twice as much for enterprise software and (unlikely) twice as much for advertising for this investment to pay off.
This excludes other sources of income, such as US government spending on the military and many “new markets” such as bio- and energy research. So, naturally, the revenue streams for AI will be much broader.
But this “Moore’s Law” idea that computing will always get cheaper is unlikely to materialize in the short term.
For comparison, the original IBM PC (IBM 5150, sold for $1565 without the hard disk) would cost about $5700 at today’s inflation prices. Today, buying a new Lenovo or Mac PC typically costs around $3,000, but don’t forget you also own an iPhone. So, as far as computing goes, over the last 45 years the “cost of computing” has gone from $5,700 to maybe $3,500. It’s not that big of a price drop.
In other words, all this great AI is very expensive, and unless it replaces a lot of other things, we’re just going to end up paying more money. And from an economic perspective, that means we need productivity, health, and other benefits we haven’t seen yet.
And companies like Oracle, Microsoft, and Workday have no intention of “replacing” their revenue with AI. growth. So are Google, Meta, SpaceX, Amazon, and Apple.
However, I am left with the idea that the price of AI is rising.
- if Corporate IT spending increases (This is clearly true), bigger ROI projects are needed, and haphazard layoffs are not the answer. Our HR 2030 project is all about finding these new ROI architectures, but we’re not there yet. And an agent who makes it easy to submit expense reports or book travel may not.
- Projects that are conceived incorrectly are scrapped. In addition to Uber, which ran out of its AI budget in a matter of weeks, Pizza Hut and Starbucks were also found to have failed AI projects, leading to $100 million in lawsuits. Therefore, after spending all that money in tokens, these projects may fail.
- Will “employment elimination” by AI be rationalized? I think so. We’re not going to buy a multi-million dollar Claude or co-pilot if it costs more than doing it manually. So we’re all going to be a little more discerning about where we focus our AI investments. I keep reading articles about blocks that exclude engineers and managers. Meta creates a 50:1 span of control. and a standard charter that excludes “low productivity” people. What would happen if the price of these “lowly productive” people became cheaper?
- Will AI replace anything else? absolutely. Over time, the money you spend on other core systems, outsourcing, consulting, accounting, and legal services all diminish. This is being discussed endlessly in the press, and service companies are being forced to redefine what they do. The service industry is 20 times the size of IT software, so there are a lot of opportunities there.
- Will AI projects become more strategic? Yeah, just like the idea of ”giving everyone a PC” didn’t happen that quickly, the idea of ”giving everyone a Claude” might stop happening too. So the widespread “adoption over value” could slow down, and I’m already seeing it happening. Many of us will end up getting a built-in agent rather than our own agent. We need to treat AI as an investment, not a random tool that everyone can play with all day long.
- Will AI agents become smarter? You have to do that. If it costs $50,000 a month to run an onboarding agency, it should be much better than a call center team that costs $25,000 a month. The trend from “assistants” to “agents” to “super agents” is accelerating. Enterprise architecture is still evolving, but more dynamic enterprises will emerge as AI is integrated into finance, HR, sales, and more. And it’s much better than an AI “sales development rep” (which doesn’t work).
- Will consumers stop paying? It’s hard to say, but if a provider decides that $20 a month isn’t enough to charge it, it will be like a cell phone where most consumers pay $75 to $150 a month just for the service. And you all know how much resistance there is to such new costs. But if I could start my own business with a Claude license for $200 a month, I would.
- Will we need a large new market? Sure. I believe companies will pay big bucks for models that serve drug discovery, energy engineering, construction engineering, and other “high-value” AI value products. In exchange for this new spending, they release even more awesome stuff and we, the consumers, end up paying for it. (Investors want returns@!)
- And finally, will the AI “market” adjust? Probably so. THigh valuations and unlimited funding for new data centers are very likely to collapse at some point, and as more people do the kind of analysis I do, more Warren Buffett types may move away from the “I’ll buy any AI stock” mindset.
Can Microsoft win the war over enterprise AI?
AI vs. the labor economy, why benefits are being cut, and the role of legacy systems.
The Context Layer (Semantic Layer) of Enterprise AI (and Where Business Rules Go)
Super agent for HR: Galileo Mars released
