Why are companies pouring trillions into AI data centers, and what's the actual expected payoff?

Started by RomanReigns26, Aug 19, 2026, 10:14 PM

Previous topic - Next topic

0 Members and 1 Guest are viewing this topic.

Topic: Why are companies pouring trillions into AI data centers, and what's the actual expected payoff?   Views(Read 125 times)

RomanReigns26

The scale of current AI infrastructure spending is genuinely staggering by almost any historical comparison, with major technology companies collectively committing well over a trillion dollars in combined capital expenditure toward data centers, specialized chips and supporting power infrastructure over just the next few years. Microsoft, Google, Amazon, Meta and a growing list of well funded AI focused companies including OpenAI and Anthropic have all announced individually massive multi year infrastructure commitments, and the specific dollar figures involved genuinely dwarf most previous eras of technology infrastructure investment, arguably rivaling entire national infrastructure programs in scale when you actually add every company's individual commitments together across the industry.

The core underlying bet driving all of this spending comes down to something researchers commonly call scaling laws, a fairly consistent observed pattern where AI model performance improves in a predictable way as you increase training data, computing power and model size together in tandem. That pattern has held up remarkably well across successive generations of models over the past several years, and companies are betting real money that it will continue holding for at least the next several generations still to come, meaning genuinely more compute translates fairly directly into meaningfully more capable models, which in turn translates into real and substantial competitive advantage for whichever company actually gets there first with the most capable system.

The expected payoff breaks down across a few genuinely distinct categories worth separating out clearly rather than treating as one single monolithic bet. The most immediate and concrete payoff is straightforward enterprise and consumer software revenue, companies selling AI powered coding assistants, customer service automation tools, productivity software and cloud AI services directly to paying business and consumer customers today, revenue streams that are already real and already growing steadily right now rather than purely speculative or hypothetical. A second, considerably larger and genuinely more speculative payoff involves labor automation at a massive economic scale, the idea that sufficiently capable AI systems could eventually perform a meaningful share of white collar knowledge work currently done by human employees, representing a market opportunity that dwarfs current AI software revenue by orders of magnitude if it actually plays out as the optimistic bull case scenario envisions.

The third and most ambitious payoff, the one most directly connected to the sheer scale of current spending, is the bet that whoever builds the most capable AI systems first captures something closer to a dominant, durable, long term competitive position, similar in spirit to how early leaders in search, mobile operating systems or cloud computing managed to build lasting moats that meaningfully compounded over many years rather than simply capturing temporary short term advantage. That's exactly why companies are willing to spend money well ahead of clearly proven, near term revenue justifying the specific dollar amount currently being spent, the strategic logic explicitly assumes that being meaningfully behind in raw underlying AI capability during this specific critical early period could translate into a difficult to ever fully overcome long term competitive disadvantage down the road.

Whether that specific bet actually pays off as cleanly as the optimistic case assumes remains genuinely uncertain and is worth treating with real skepticism, and this is exactly where serious informed disagreement among knowledgeable analysts and researchers actually lives. Skeptics reasonably point out that scaling laws could plausibly hit meaningful diminishing returns well before the truly transformative capability improvements the optimistic bull case explicitly depends on ever actually materialize in practice, that current AI revenue still represents a small fraction of the enormous capital being spent upfront right now, and that several previous major technology infrastructure buildouts throughout history, fiber optic cable during the dot com era being a frequently cited historical example, ultimately did prove broadly useful in the long run but only after a painful period of real financial overbuilding, genuine investor losses, and substantial corporate consolidation among the initial wave of companies that spent aggressively and early

Lantern

The dot com fiber optic comparison is honestly the single most useful historical parallel for thinking clearly about this whole situation, since it captures both sides of the actual argument simultaneously in one example. All that fiber genuinely did eventually prove enormously useful once real demand actually caught up to the built capacity years later, but a huge number of the individual companies that built it aggressively and early went bankrupt or got acquired for pennies on the dollar well before that broader payoff genuinely materialized for the industry as a whole.

Margin

The competitive moat logic honestly makes a lot of intuitive strategic sense to me when I think it through carefully, being meaningfully behind during a critical formative early period in a new and rapidly evolving technology category really can translate into a lasting structural disadvantage that's incredibly difficult to ever fully close later. Whether AI capability specifically ends up working that particular way in practice remains a real and completely open question though, not every past technology category has actually played out following that exact same specific historical pattern.
Opinions are my own. Obviously.

Terry_33

Current AI revenue representing just a small fraction of total capital being spent upfront right now is honestly the statistic that should give any reasonable person real pause here, regardless of how compelling the longer term strategic bull case narrative ultimately sounds on paper. Plenty of previous technology bubbles throughout history featured that exact same specific gap between current actual revenue and current capital spending, and not every single one of those historical bubbles eventually justified itself after the fact the way the fiber optic buildout example genuinely did.

Jackson_23

Worth genuinely emphasizing just how much power infrastructure specifically factors into this whole spending picture beyond just the chips and servers themselves, since data centers at this particular scale are running directly into real hard physical grid capacity constraints in a lot of regions right now. That's a meaningfully different and considerably harder kind of bottleneck to solve quickly compared to simply manufacturing more chips faster, since power grid buildout timelines are measured in years rather than months in most cases.

WWEGary20

Scaling laws holding up consistently so far is genuinely true as an empirical historical observation, but past pattern reliably continuing indefinitely into the future is exactly the actual bet being made here with all this capital, and it's a real and legitimate open empirical question rather than a settled established fact anyone should simply assume will keep holding forever without limit.

AJStyles04

The white collar labor automation payoff scenario feels like the genuinely load bearing assumption underneath most of the most aggressive current infrastructure spending decisions being made across the industry right now. If that particular scenario doesn't actually materialize at meaningful real world scale within a reasonable timeframe, or takes considerably longer to materialize than current spending plans clearly assume, a whole lot of currently committed capital is going to end up looking genuinely premature and poorly timed in hindsight several years from now.

Bayley_US

Breaking the payoff down into those three genuinely distinct categories is honestly really clarifying and I wish more mainstream coverage of this spending actually did that instead of just treating it as one single undifferentiated giant bet on AI broadly. Current enterprise software revenue is genuinely real today, the labor automation case is much more speculative and uncertain, and the durable competitive moat argument is really more of a strategic defensive positioning bet than any kind of concrete near term revenue projection at all.

Current

A tangent worth adding is that AI infrastructure could change the economics of software development itself. If coding becomes cheaper and faster, more organisations may build software that previously wasn't worth developing.

That creates a feedback loop. Cheaper development increases the number of applications, which creates more demand for compute, which justifies more infrastructure, which can reduce compute costs further.

The same could happen with smaller businesses. A company that couldn't afford a custom analytics system might use AI tools to build one from existing data without hiring a large specialist team.

But feedback loops can work in the other direction too. If AI tools become commodities and prices collapse, infrastructure providers may struggle to capture much of the value even while usage explodes.

That distinction between economic value and captured revenue is really important. Society can become much more productive without every company supplying the underlying technology earning extraordinary margins.

We've seen this with many computing technologies. The capability becomes cheap and ubiquitous, while the real economic benefit spreads across thousands of industries.

If AI follows that pattern, the trillion-dollar spending could still be economically significant even if some individual infrastructure investments don't produce spectacular financial returns.

Terry_33

One final thought: the scale of the spending should not make us assume the technology has to produce an equally huge pile of direct profits to be useful. Electricity grids, roads and communications networks create economic value far beyond the revenue collected by whoever owns the infrastructure.

If AI becomes a general-purpose productivity technology, the gains could show up in better products, faster research, lower operating costs and new businesses that don't exist yet.

That is also why forecasts are so difficult. The most valuable applications may be the ones nobody has thought of today. Giving developers abundant compute can create experimentation that eventually produces something much more valuable than the original use case.

Of course, that argument can be abused to justify any amount of spending. "We don't know what will be invented" isn't a blank cheque. Capital still has an opportunity cost, and infrastructure has to earn enough to cover construction, hardware, energy and financing.

So there needs to be a reality check alongside the enthusiasm. Watch utilisation, pricing, margins, energy efficiency, customer retention and measurable productivity rather than just counting data-centre announcements.

If those indicators keep improving, the huge buildout starts looking more defensible. If they don't, the industry may eventually discover that building compute was easier than finding profitable things to do with all of it.

That uncertainty is probably the most interesting part of the whole debate. The money is being spent now, but the real verdict will come from what people actually do with the machines once they are built.

Wolfhound39

The comparison I'd make is with the buildout of previous infrastructure booms. Nobody spends billions on a railway because each individual passenger is guaranteed to be profitable. They spend because the network can support many different economic activities once it exists.

AI data centres could work similarly. A single facility might support model training, inference, scientific computing, enterprise workloads and rented cloud capacity. The economics become more interesting when the same physical infrastructure can serve multiple customers and use cases.

The tricky bit is depreciation. Chips don't necessarily remain the most competitive hardware forever. If a new accelerator generation provides much better performance per pound, older equipment can lose economic value even though it still works perfectly well.

That creates a strange situation where a data centre can be physically useful while some of its most expensive equipment is becoming less attractive. Traditional infrastructure investors aren't necessarily used to that rate of technological change.

On the other hand, software improvements could reduce the amount of compute required for a given task. Better algorithms and smaller models might mean you can do more work with the same hardware.

That cuts both ways. Efficiency improvements could reduce demand for raw compute, but they could also make AI cheap enough to deploy in vastly more places. More efficient computing doesn't automatically mean less total computing.

That rebound effect is probably one of the most important pieces of the puzzle. If AI becomes ten times cheaper to run and people consequently use it twenty times more, the infrastructure market can still grow substantially.

SpikeDudley07

The electricity side is where the numbers become particularly interesting to me. A data centre isn't just a warehouse full of GPUs. It is also a long-term commitment to power, cooling and grid capacity.

That means the infrastructure investment can create value beyond the AI company itself. Utilities, construction firms, networking suppliers and energy developers can all participate in the buildout.

But it also creates a genuine risk if capacity gets built faster than useful demand. Electricity infrastructure has much longer planning horizons than software, so getting the timing wrong isn't something you fix with a quick patch on a Tuesday afternoon.

There is a similar issue with cooling. High-density computing produces enormous amounts of heat, and moving from one generation of hardware to another can change the cooling requirements significantly.

This makes AI infrastructure partly an industrial story rather than just a technology story. Concrete, transformers, power distribution, fibre, cooling equipment and physical land all matter.

That may actually be one reason the investment looks so huge compared with earlier software cycles. You are not just paying programmers to build an application. You are constructing a physical industrial base for computing.

Whether that base becomes highly profitable is still an open question, but the scale of the physical commitment is easier to understand once you include all those layers.
404: Signature not found

Sandworm

The easiest way to understand the spending is to separate revenue from strategic positioning. A company may not know exactly how much money a particular AI service will make in five years, but it can still believe that having insufficient compute would leave it at a serious disadvantage.

Think about a retailer deciding whether to build ten warehouses before knowing exactly how many orders it will receive. That sounds reckless until you remember that the warehouses themselves can determine how quickly the company can fulfil orders. AI infrastructure has a similar capacity argument behind it.

There is also a very mundane reason for the huge numbers: data centres are expensive physical assets. You need buildings, electricity, cooling, networking, chips, backup systems and all the supporting equipment. Once you multiply that across thousands of machines, the capital requirement gets enormous very quickly.

The payoff doesn't necessarily have to come from selling chatbot subscriptions. It can show up through advertising, cloud services, enterprise software, developer tools, search, productivity products or lower costs for existing businesses.

Another possibility is that some of the infrastructure becomes a platform other companies rent. That changes the economics because the owner isn't relying entirely on one consumer application becoming wildly profitable.

Where I get cautious is assuming utilisation will automatically remain high. A data centre full of expensive accelerators only produces a good return if customers are willing to pay enough to keep those resources busy.

So the real question isn't simply whether AI will be valuable. It is whether the value created per unit of compute will grow quickly enough to justify the enormous capital being committed. That is a much harder question.

CosmicRay65

One thing I would add to the discussion is that inference could eventually be embedded into products people don't even think of as AI products. If a camera, search engine, accounting package or industrial system quietly uses a model in the background, the amount of compute consumed can become enormous without users ever visiting an AI website

That is where the scale could really surprise people. Ten million people paying for a chatbot is one workload. Hundreds of millions of devices and business processes making smaller AI calls throughout the day is a completely different workload

It also means the infrastructure has to become cheaper and more efficient. A company can't put an expensive large model behind every tiny feature and expect customers to accept unlimited costs

That creates pressure for model routing, caching, smaller models and specialised hardware. Not every request needs the biggest possible system

So I wouldn't expect the future to be a simple story of ever-larger models consuming ever-larger amounts of compute. There will probably be a messy stack of tiny, medium and huge models doing different jobs

That could make the infrastructure market much broader than the current training-focused picture suggests

The joke is that after all these trillions are spent, half the AI running our lives might be some tiny model quietly deciding whether a notification needs to be shown

Aurora Kai

The productivity argument is probably the most important one outside the technology industry. If AI can make a worker 20% more productive at certain tasks, the economic impact can be enormous even if the AI provider charges relatively little for the service

Take a software developer as a simple example. If an assistant helps with routine code, documentation and debugging, the company doesn't necessarily need to fire people. The same team might simply complete more projects in a year

That distinction matters because economic value doesn't always appear as direct revenue. More output from the same workforce is valuable even when headcount remains unchanged

There is a caveat, though: productivity improvements in controlled demonstrations don't automatically translate into workplace gains. People still need to review outputs, integrate tools into existing systems and deal with errors

A tool that generates a draft in thirty seconds but requires twenty minutes of checking hasn't necessarily saved much time. The quality and reliability of the workflow matter more than the raw generation speed

That is why I think the next phase will be less about impressive model benchmarks and more about boring operational measurements. How many hours did it save? How often was human correction needed? Did customer satisfaction change? Did the process actually cost less?

Those answers will tell us much more about whether trillions in infrastructure spending can translate into lasting economic value
Making the internet slightly better one post at a time

CipherMonk37

The biggest potential payoff may be boring old automation rather than replacing entire professions. If AI can handle repetitive administrative work, the value can accumulate across millions of small tasks

Take insurance claims, invoices or internal company documents. A system that extracts information, checks it against rules and prepares a draft for a human can save minutes on each case. Multiply those minutes across millions of cases and the economics become substantial

The important word there is assist. Many workflows can tolerate an AI making a draft that a person checks. They may not tolerate an autonomous system making an irreversible decision without oversight

That distinction also affects infrastructure requirements. A high-volume workflow may favour a smaller, cheaper model because the goal is reliable extraction rather than dazzling reasoning

So the future compute demand might not be dominated by giant frontier models answering every question. A lot of it could come from specialised systems doing narrow jobs at enormous scale

That is a much more prosaic vision of the AI economy, but probably a useful one. Most successful technology ends up being embedded in ordinary processes rather than permanently presented as a futuristic novelty

If that happens here, the payoff from all this infrastructure could be spread across thousands of mundane efficiency gains rather than one giant killer application

Cheetah

The cloud model is probably the part that makes the numbers easier to digest. A company doesn't necessarily need to sell an AI product to every person on the planet if it can rent compute to thousands of businesses

Think of it like a hotel. The owner doesn't care whether every guest is there for the same reason; the business works because rooms are rented repeatedly. A cloud platform can similarly sell capacity to companies with completely different applications

Of course, compute isn't a hotel room. Hardware depreciates, energy costs money and workloads fluctuate, so utilisation is critical. An empty accelerator is basically an expensive sculpture

That makes scheduling and workload management surprisingly important. If training jobs can use capacity at one time and inference workloads use it at another, the same infrastructure can generate more revenue

There is also potential for different quality tiers. Not every AI task needs the newest, fastest accelerator. Older hardware can still be useful for less demanding workloads if the economics are right

That gives providers another way to extend the useful life of infrastructure. The most powerful hardware can serve demanding applications while older machines handle cheaper workloads

So the data-centre business may eventually look less like buying one giant computer and more like operating a constantly optimised fleet of different computing resources

BretHart

A lot of the expected payoff seems to be tied to inference rather than training. Training a giant model gets the headlines, but once a useful model exists, every customer interaction can generate ongoing compute demand

Imagine an enterprise using AI for customer support. The model isn't trained from scratch every time someone asks about an invoice, but millions of those interactions still require processing. Multiply that across search, office software, coding assistants, customer service and other applications and the recurring workload becomes much larger

That also explains why latency matters. A company may be perfectly happy to wait hours for a training job, but a consumer asking a question expects an answer almost immediately. That pushes providers toward having compute capacity distributed and available when demand arrives

The economics therefore aren't just about maximum performance. Cost per useful response, energy consumption, networking and utilisation all matter

There is a counterargument, though. If models become dramatically more efficient, today's assumptions about how much inference capacity is needed could look excessive. Nobody can know precisely how much efficiency will arrive alongside demand growth

So I wouldn't interpret every infrastructure announcement as proof that the spending will pay off. It is a bet on both sides of the equation: more demand for AI services and continued improvement in the economics of serving them

If either side disappoints badly, returns could look very different from the optimistic scenarios

Henry10

There is a strange historical irony here. Data centres used to be something most people never thought about, and now they are becoming a central part of infrastructure planning because AI has made compute demand so visible

That could have consequences for regions trying to attract investment. A new facility can bring construction work and infrastructure upgrades, but it can also require substantial electricity and other resources

The local economics therefore deserve attention. A community might welcome the investment while still asking how many permanent jobs it creates, how much power it consumes and what public infrastructure is needed to support it

From the company's perspective, location is also about more than cheap land. Access to power, network connectivity, cooling conditions, permitting and reliability can all affect operating costs

So the phrase "AI data centre" makes it sound like a purely digital asset when it is actually a very physical one. You need roads, substations, fibre and buildings before a single model request can be processed

That makes the investment cycle somewhat different from the classic software boom. Software can scale almost instantly once it is written; physical compute capacity has to be planned and constructed

Maybe that is why companies are spending so aggressively now. They are trying to secure the physical foundations before demand reaches the point where capacity itself becomes the bottleneck
Here more than I should be

Amelia82

What makes me cautious is the phrase "expected payoff" because there probably isn't one single expected payoff. Different investors are buying different things

A chip company wants to sell hardware. A cloud provider wants to rent compute. A model company wants to sell access to intelligence. A software company wants to increase the value of its existing products. An enterprise wants lower costs or higher output

Those businesses can all benefit from the same infrastructure while measuring success completely differently

For a cloud provider, utilisation and revenue per accelerator might matter most. For an enterprise, the important figure might be cost per resolved support ticket or hours saved by employees

That is why giant spending totals can be misleading when treated as one giant investment pot with one giant return. The capital is flowing through a chain of businesses, and each layer has its own economics

Some parts of the chain may earn excellent returns while others struggle. That's normal in a technology ecosystem

I'd be much more interested in seeing infrastructure spending broken down by who is funding it, who owns the assets, what workloads they expect to run and how those workloads are supposed to generate cash. That would tell us far more than another headline about the total number
Coach says my form's good, my code says otherwise

Neil82

The financial question I'd ask is not "will AI make money?" because obviously some AI businesses already do. The better question is whether returns on the incremental infrastructure remain attractive as spending gets larger

The first billion dollars invested in a useful capability can have a very different return profile from the next hundred billion. Early capacity might serve the easiest and most valuable workloads, while later capacity has to chase lower-margin uses

That is where competition matters. If several companies build enormous amounts of compute at the same time, customers may gain bargaining power and prices could fall

Falling prices aren't necessarily bad for the technology. They could accelerate adoption dramatically. But they make the infrastructure owners' return calculations more complicated because the same hardware may generate less revenue per unit of capacity

There is also financing risk. Huge infrastructure projects need capital, and investors eventually want evidence that the assets will generate sufficient cash flow. A strong technology story does not automatically guarantee strong returns on every individual investment

So I suspect there will be winners even if some of the spending proves excessive. The difficult part is identifying which layer captures the value: chipmakers, cloud providers, model companies, software vendors, customers, or some combination

That distribution of value is probably more interesting than the headline spending figure itself

OffsideTrap

There is a less glamorous answer too: companies are spending because compute availability itself has become a competitive advantage. If your engineers have to queue for scarce hardware while another company can run experiments immediately, that affects development speed

For an AI lab, the difference between waiting days for a large training run and starting it immediately can compound across hundreds of experiments. Faster iteration can be valuable even before a new product reaches customers

That makes compute something like factory capacity. A manufacturer doesn't only buy machines because each machine directly earns money; it buys them because production capacity determines what the business can make and how quickly it can respond

The catch is that excess factory capacity is expensive. If the expected demand doesn't arrive, the company is stuck paying for assets that aren't being used efficiently

AI infrastructure therefore involves a classic capacity-planning problem, just at an unusually large scale. Build too little and you constrain growth; build too much and your capital sits idle

I suspect the companies involved are deliberately accepting some of that risk because being short of capacity at a critical moment could be more damaging than carrying some spare capacity

Whether that strategy works financially is something we'll only really know after the infrastructure has been operating for years, not months

Theo50

One thing that gets lost in the trillions discussion is that not all spending is necessarily coming from the same economic pocket. A cloud provider building a data centre for external customers has a different business model from a company building capacity mainly for its own products

The first can make money by selling compute directly. The second might use the infrastructure to improve advertising, search, office software or other services where the AI investment is buried inside a much larger revenue stream

That makes the payoff surprisingly difficult to measure. Suppose an AI assistant makes an office suite more useful and reduces customer churn. The financial benefit may never appear as a neat line labelled "AI revenue"

There are also productivity gains inside companies themselves. If programmers, analysts, support workers and researchers can complete tasks faster, the benefit may show up as lower labour costs or greater output rather than additional sales

That doesn't mean every claimed productivity gain will materialise. Measuring whether an AI tool genuinely saves time is harder than putting a chatbot in front of employees and announcing that it exists

A useful metric would be something like cost per completed task rather than cost per token or number of AI users. If an AI system costs £1 million but replaces work worth £3 million, that is a very different proposition from a system that simply generates lots of impressive demos

The business case ultimately has to reach that level of practical measurement
Thanks to everyone who has ever answered a question here. You're the real model.

Estuary59

The strongest counterargument to all this spending is simply that hardware becomes obsolete. If today's accelerators are dramatically less efficient than tomorrow's, locking in huge quantities of current equipment could turn into a problem

We've seen versions of this before in computing. A machine can be perfectly functional but economically unattractive because a newer generation can do the same work for much less money or energy

AI makes that particularly interesting because software and hardware are improving together. Better models may require less compute, while better chips can provide more compute per watt. Either development can change the amount of infrastructure needed

On the other hand, efficiency improvements often increase usage. If running an AI assistant becomes ten times cheaper, companies may put assistants into ten times as many workflows. The savings per request can therefore coexist with rising total demand

This is why I wouldn't assume efficiency automatically destroys the investment case. It could actually expand the market

The real uncertainty is the balance between those effects. Nobody has a perfect model for how much additional demand appears when AI gets cheaper

That's probably why the spending looks so extreme: companies are effectively placing long-term bets on several uncertain variables at once

Exposed Thea

There is a bit of a chicken-and-egg problem here. Companies need massive compute capacity to build and improve better AI systems, but they also need better AI systems to justify massive compute capacity

If everyone waits for proven demand before building infrastructure, capacity may arrive too late to support the next wave of applications. That encourages companies to invest ahead of demand

The danger is obvious: everyone can convince themselves that future demand will be enormous, and then discover that the market isn't quite as large as expected. Technology history has plenty of examples of infrastructure being built on optimistic assumptions

Still, AI has one unusual feature compared with some previous technology bubbles: there are already large numbers of people paying for AI services. The question is more about the scale and durability of that demand than whether demand exists at all

Enterprise customers are especially important because a company may pay thousands or millions for software if it produces measurable savings. Consumer subscriptions are useful, but enterprise adoption could provide a much larger pool of revenue per customer

The strongest business cases I've seen tend to involve boring workflows rather than flashy demos. Automating document processing, software testing, customer support or internal search may not look revolutionary, but those are areas where savings can be measured

If that pattern continues, the data-centre spending becomes easier to justify. If businesses mostly use AI for novelty and experimentation, the maths gets much harder

RicFlair

There is another payoff that rarely gets mentioned: strategic control. If a company relies entirely on someone else's compute, it may be vulnerable to price increases, capacity shortages or changes in access terms

Owning or controlling infrastructure can therefore have value even when the accounting return isn't immediately spectacular. It gives the company more control over availability, model development and product launches

That doesn't make every internal data centre economically sensible, but it explains why a company might accept a lower short-term return than a pure infrastructure investor would

The same logic applies at a national level, although that gets into a much broader debate. Compute capacity can be viewed as part of technological infrastructure, much like networks or semiconductor manufacturing

There is a risk in interpreting strategic spending as proof of commercial profitability, though. A company can rationally spend money for defensive reasons even when the investment itself isn't expected to produce a huge direct return

So when someone says "they are spending trillions, therefore they must know AI will generate trillions," I wouldn't follow that logic. Corporate investment can reflect fear of being left behind as much as confidence about future profits

Sometimes the payoff from an investment is simply having the option to participate in whatever market emerges next

Stag

There is a good reason to be both impressed and sceptical about the spending. The physical buildout is clearly enormous, but the revenue assumptions behind it still depend on how useful AI becomes in real workflows

A flashy demo can justify excitement, but businesses ultimately have to ask whether the result is better, cheaper or faster than the alternative. Sometimes the answer will be yes. Sometimes a traditional database, search system or ordinary piece of software will still be better

That is why I like the three-part way of looking at the payoff: direct AI revenue, productivity gains and strategic value. They are all real categories, but they shouldn't be mixed together when evaluating a particular investment

Direct revenue is relatively easy to understand. Productivity gains are harder to measure but potentially huge. Strategic value is the hardest to put into a spreadsheet because it can involve maintaining competitive options rather than generating immediate cash

There is also a timing issue. A company might spend heavily today while the largest productivity benefits arrive years later. That doesn't automatically make the investment irrational, but it does mean investors have to tolerate a long period where the spending runs ahead of obvious returns

Personally, the part I find most convincing is the possibility that cheap AI becomes a general-purpose utility used throughout business. The part I find least convincing is any claim that we already know exactly which applications will justify today's capacity

The infrastructure may be the foundation of something enormous, but foundations don't tell you what building will eventually stand on top of them

Joanne_24

You are spot on with the auction analogy, except it feels less like art and more like the world's most aggressive game of musical chairs. Every big tech giant knows the music is going to stop eventually, but nobody wants to be the one standing without a chair when the power bill comes due. The ROI math they are throwing around right now relies on literally every human being on the planet subscribing to fifteen different chat bots just to justify the electricity alone

Meanwhile, my local power company can't even keep the lights on during a moderate summer breeze, yet we are apparently supposed to power an entire artificial god out of a converted shed in rural Oregon. Call me cynical, but I reckon the real breakthrough won't be AGI, it will be finding a way to declare bankruptcy without anyone noticing.

Related Topics (6)