Everyone Is Betting Trillions on AI. So Why Has Nobody Found the Business Case?

The AI industry is burning money at a scale that makes the dot-com era look like a small campfire. Valuations are hitting the stratosphere. Models are dropping every few weeks. And yet, when you ask most businesses what AI has actually done for their bottom line, the honest answer is: not much. Here is the full picture of what is really going on, and what it means for Kenya and the rest of the world watching from the outside.

Let us start with a number that should give everyone pause.

By 2028, OpenAI is projected to post a single-year operating loss of $85 billion. Not cumulative losses across several years. A single year. And the situation does not improve from there: the company is not expected to break even until 2030, by which point it will have burned through roughly $143 billion in total cash outflows. This is a company currently valued at over $850 billion that has never made a profit in its life and, according to its own financial projections, will not make one for at least four more years.

Meanwhile, Anthropic, the company building the AI model you may have used to write your last work email, saw its annualized revenue jump from $381 million in 2024 to $30 billion by April 2026. That trajectory sounds incredible until you realize the company is still losing money and burning through capital at a rate that requires it to raise billions every few months just to keep the lights on. Its IPO, expected as early as October 2026, is targeting a valuation somewhere between $380 billion and $900 billion depending on which investor you ask.

Both companies are losing money. Both companies are worth hundreds of billions of dollars. Neither of them has a clear path to the kind of profit margins that would justify those valuations for years to come.

And somewhere in the background, Nvidia quietly posted record revenue of $57 billion in a single quarter, with gross margins sitting comfortably above 73%.

If you are confused about what is happening here, you are paying attention.

The Numbers Behind the Hype

To understand the AI economy as it actually exists today, you need to separate three things that the industry has deliberately blurred together: revenue, growth, and profit.

Revenue is surging across the board. OpenAI went from $3.7 billion in revenue in 2024 to a $24 billion annualized run rate by April 2026. Anthropic went from $381 million in 2024 to a $30 billion annualized run rate in roughly the same period. These are genuinely remarkable growth rates. No traditional business has ever scaled this fast.

But growth is not profit. Every dollar these companies earn is being outpaced by what they spend to earn it. OpenAI's biggest cost is the computing infrastructure it leases from Microsoft to run its models. Anthropic's biggest cost is compute from Google and AWS. The fundamental economics of the AI model business are brutal: training frontier models costs billions of dollars, inference (the act of responding to a user query) costs significant money per interaction, and the competitive pressure to keep releasing more capable models means the training cost never stops. You do not train a model once. You train it, then you train a better one, then another, then another. Each one costs more than the last.

This is why we are getting a new major model release roughly every three weeks across the industry. It is not just technical ambition. It is a structural necessity. If you stop releasing better models, your customers migrate to whoever is releasing better ones. The treadmill never stops, and the electricity bill for the treadmill keeps going up.

The entity that benefits cleanest from all of this is Nvidia, which is selling the hardware that powers every single model from every single company. Nvidia's data center revenue grew 66% year-on-year to $51.2 billion in a single quarter, driven almost entirely by AI demand. Jensen Huang, Nvidia's CEO, described the situation with characteristic precision when he said "Blackwell sales are off the charts, and cloud GPUs are sold out." The picks-and-shovels metaphor from the California Gold Rush has never been more appropriate. The people rushing for gold may or may not get rich. The person selling the shovels is already wealthy.

In Kenya and across Africa, this dynamic is being watched from a distance, but it is not entirely abstract. The Nairobi AI Forum 2026, held in February, brought together governments, private sector leaders, and tech innovators to chart pathways for AI adoption. At that forum, the African Development Bank and the UNDP launched the AI 10 Billion Initiative, a partnership targeting $10 billion in mobilized capital by 2035 to unlock what they project could be 40 million new jobs across the continent. AI is projected to contribute up to $2.4 billion to Kenya's GDP by 2030. The question worth asking is: where does that value come from, and who captures it, if the companies building the technology cannot yet capture it themselves?

The Amazon Comparison Everyone Gets Wrong

When confronted with the profitability question, defenders of AI economics almost always reach for the same historical analogy: Amazon. Amazon lost money for years, they say. Amazon was patient. Amazon built the infrastructure, and eventually the profits came. Be patient with AI.

The comparison is not wrong. But it is importantly incomplete.

Between 1994 and 2003, Amazon reported net losses in nine of its first ten fiscal years. Its largest single-year loss was $1.4 billion, during the dot-com bubble in 2000. Across its first 17 quarters as a public company, Amazon lost a combined $2.8 billion. Those are real numbers, and the patience of investors who held through that period was eventually rewarded with one of the greatest wealth creation stories in business history.

Now consider that OpenAI alone is projected to lose $85 billion in a single year by 2028. That is more than 60 times Amazon's worst-ever annual loss, happening in one year, at a company that is less than a decade old. The scale of the spending is not comparable. It is a different category of financial commitment entirely.

More importantly, Amazon's losses were funding something with provable, growing utility. Every dollar Amazon lost was building warehouses, logistics networks, and eventually AWS, all of which had clear and direct paths to revenue. Customers were already using the products and paying for them. The losses reflected the cost of scaling what was working, not the cost of figuring out whether it worked.

The honest question about AI is whether we have found "what is working" yet, at the scale required to justify the spending. The evidence is, to put it gently, mixed.

The FOMO Factory: Why Everyone Keeps Spending

Here is a statistic that deserves more attention than it gets.

According to MIT's GenAI Divide study, which analyzed 300 public AI deployments alongside executive surveys, 95% of generative AI enterprise pilots failed to deliver measurable financial returns within six months. S&P Global found that 42% of companies scrapped most of their AI initiatives in 2025, up from just 17% the year before. IBM's research found that only 25% of AI initiatives deliver expected ROI, and only 16% have scaled enterprise-wide. Morgan Stanley found that just 21% of S&P 500 companies could cite a measurable AI benefit at all.

These are not fringe numbers from skeptical researchers. These are findings from the major consulting firms and banks whose clients are the companies spending billions on AI.

And yet, from the same surveys, investment is not slowing down. It is accelerating. Average enterprise AI spending is projected to jump 65% from about $7 million per company in 2025 to $11.6 million in 2026, even as most firms cannot demonstrate a return. A Deloitte survey of 1,854 executives captured the psychology behind this contradiction in one quote from an anonymous telecommunications executive: "Everyone is asking their organisation to adopt AI, even if they don't know what the output is. There is so much hype that I think companies are expecting it to just magically solve everything."

This is the FOMO dynamic in its purest form. No board wants to be the board that told management not to invest in AI and was later blamed for falling behind. No CEO wants to face investor questions about why their company has no AI strategy. The adoption is driven as much by the fear of being seen as a laggard as by any concrete analysis of what AI will actually do for the business.

The result is a peculiar economic loop. Companies spend on AI tools because competitors are spending on AI tools. AI companies use that spending to justify continued investment in model development. Model development requires more compute. More compute requires more Nvidia GPUs. Nvidia's profits fund its own continued dominance of the hardware stack. And the companies actually spending on the tools are still largely waiting for the payoff.

In Kenya, this dynamic plays out with an added layer of complexity. A Zoho report on Kenyan AI adoption found that businesses are prioritizing AI investment in customer service and software development, and that Kenya is emerging as a model for "youth-driven, privacy-conscious AI adoption." These are real and encouraging signals. But the ROI problem that is plaguing enterprises in Europe and North America does not disappear just because you are operating in a developing market. If anything, the margin for error is smaller. A company in Nairobi that spends aggressively on AI tools that deliver no measurable return has less cushion to absorb that loss than a Fortune 500 company doing the same thing.

So Who Actually Makes Money from AI?

It helps to think about AI's economic structure the way we think about the internet's economic structure in the 1990s and early 2000s.

When the internet was being built, the companies laying the fiber and building the routing infrastructure were not the ones that eventually made the most money. Many of them went bankrupt in the dot-com bust. The companies that made generational wealth were the ones that built applications on top of that infrastructure: Google on top of the web, Facebook on top of social connectivity, Amazon on top of e-commerce, Netflix on top of broadband. The infrastructure providers enabled the value. The application layer captured it.

There is a strong argument that AI's economic structure will resolve similarly. The model companies, OpenAI, Anthropic, Google DeepMind, are building the infrastructure. They are training and maintaining the foundational models. They are absorbing the compute costs, the data licensing costs, the safety research costs. They are the fiber-layers. Whether they ever fully capture the value of what they are building is genuinely uncertain.

The application layer, companies that build specific, high-value tools on top of these models, is where the cleaner business cases currently exist. GitHub Copilot, which charges developers for AI-assisted coding, has achieved real enterprise stickiness. Specialized AI tools in legal research, medical diagnosis assistance, and financial analysis are finding specific workflows where the productivity gain is measurable and the willingness to pay is high. The key in all these cases is the same: a narrow, well-defined problem where AI's capability matches the need precisely, rather than a vague mandate to "use AI."

This is where Africa's opportunity sits, and it is a real one. The continent does not need to compete with OpenAI or Anthropic at the model layer. It almost certainly cannot, given the compute infrastructure requirements. What it can do is build at the application layer, deploying these models against problems that are specific to the African context and where no equivalent solution currently exists. Agricultural AI tools that help smallholder farmers in Western Kenya optimize planting decisions based on hyperlocal weather and soil data. Healthcare diagnostic tools that extend the reach of doctors in areas where the doctor-to-patient ratio is critically low. Financial services tools that deepen credit access in a market where M-Pesa has already demonstrated the extraordinary demand for mobile-first financial infrastructure. These are not generic AI use cases. They are specific, high-value problems where a well-built application layer product could generate sustainable revenue.

The $10 billion AI initiative announced at the Nairobi Forum is significant precisely because it is focused on building the foundations for this kind of work: data infrastructure, compute access, skills development, and policy frameworks. The 130 African innovators who received GPU compute access as part of that forum's announcements are the application layer in embryonic form.

The Google Dilemma: Cannibalizing Your Own Success

No company illustrates the AI profitability paradox more starkly than Google, and we have written about this in detail before. But the economic dimension of what Google is doing deserves a closer look in this context.

Google's traditional search business is, by most metrics, the most efficient advertising engine ever built by a private company. A user types a query, Google returns links, the user clicks through to a publisher's website, the publisher shows ads and earns revenue, and Google's AdSense takes a cut. The entire ecosystem runs on a single click. It is a near-perfect flywheel.

AI Mode, Google's new default search experience, breaks that flywheel deliberately. When Google generates a full AI answer to your query, you do not click through to a publisher. You read the answer, get what you need, and leave. Google earns nothing from that interaction. The publisher earns nothing. The AdSense ecosystem earns nothing. Research tracking 68,000 real searches found that when an AI summary appeared, users clicked on results only 8% of the time, compared to 15% without one. Zero-click searches jumped from 56% in 2024 to 69% by May 2025.

Google is spending billions of dollars to build a product that, in its current form, is less profitable than the product it replaces. The logic is defensive rather than commercial: if Google does not do this, ChatGPT and Perplexity will take its users anyway. Better to cannibalize your own revenue than let someone else do it.

This is the innovator's dilemma playing out in the most dramatic and expensive way possible. And it points to something important about the AI economy more broadly. Much of the investment happening right now is not being made because someone has found a better business model. It is being made because the alternative, standing still while competitors move, feels more dangerous than spending.

The question nobody can fully answer yet is what the replacement business model for AI-powered search actually looks like. Google has not found it. The publishers losing traffic have not found a substitute for it. The entire digital advertising ecosystem that has funded most of the open internet for two decades is being restructured in real time, and the destination is not yet visible.

For Kenyan publishers and digital media businesses, this is not a distant problem. Kenya's internet advertising market is growing at 16% compound annual growth, the fastest rate globally according to PwC's projections, heading toward $470 million by 2029. That growth is happening on a foundation that AI-powered search is actively undermining. The traffic model that Kenyan digital publishers have built their businesses around is the exact model that AI Mode is designed to make obsolete. The fastest-growing advertising market in the world is running toward a cliff that has already appeared in more mature markets.

The Valuation Paradox: Pricing a Future That Hasn't Arrived

Let us address the question that sits at the center of all of this directly. How do you justify valuing a loss-making company at $850 billion?

The answer is that you are not valuing the company as it exists today. You are pricing in a specific bet about the future: that AI will be genuinely transformative at civilizational scale, and that whoever wins the AI race will capture an extraordinary share of that value. The market is essentially saying: if AI does what its proponents claim it will do, the company that ends up as the dominant platform for that technology will be worth more than any company that has ever existed. The valuation reflects the probability-weighted expected value of that outcome.

This is not irrational. It is exactly how investors priced Amazon in the late 1990s, when the company was burning cash and the conventional wisdom was that it would never survive. A $1,000 investment in Amazon at its dot-com-era peak, even after the 90% crash, would be worth over $15,000 today. The investors who understood what Amazon was building and held through the losses were right.

But the dot-com era also produced Pets.com, Webvan, Kozmo, and hundreds of other companies with similar valuations and similar promises that are now gone. For every Amazon, there were dozens of companies that promised to reshape the world, burned through investor capital at spectacular rates, and then simply ceased to exist. The pattern of high valuations on pre-profit technology companies does not predict success. It predicts a sorting: some companies will deliver on the promise, and their investors will be rewarded enormously. Many will not, and their investors will lose everything.

The uncomfortable truth about the current AI landscape is that we are still in the phase where it is genuinely difficult to tell which category most of these companies fall into. OpenAI has the largest consumer AI product in history in ChatGPT, with over 800 million weekly active users. Anthropic has overtaken it in revenue despite having roughly 5% of ChatGPT's consumer user base, which tells you something important about where the real money in AI is coming from: enterprise contracts, not individual subscriptions. Google has the most complete AI stack of any company in the world, from chips to models to distribution. But none of them have yet demonstrated the kind of durable, scalable profit that would justify their current valuations based on fundamentals rather than future-state expectations.

Anthropic's path to profitability looks cleaner than OpenAI's. Its cash burn is projected to drop to 9% of revenue by 2027, with profitability expected by 2028. The enterprise focus, prioritizing high-value business contracts over consumer scale, produces better unit economics than a platform built primarily around free consumer access. But "better than OpenAI" and "justified by a $500 billion IPO valuation" are not the same sentence.

What Has to Be True for This to Pay Off

There is a version of this story that has a clean, profitable ending. It requires several things to be true simultaneously.

First, AI has to find its "killer app." The internet's killer app was email, then search, then social networking. Each of these unlocked a business model that could scale. AI's most credible candidate right now is coding assistance. GitHub Copilot and similar tools have achieved real productivity gains for developers, and developers are willing to pay for them. But coding is a relatively narrow market. For AI to justify its current investment levels, it needs a killer app that serves the mass market the way search did.

Second, inference costs have to fall dramatically. The cost of running AI models, as opposed to training them, is still high enough that many potential use cases are not commercially viable at scale. If inference costs follow the same trajectory that storage and bandwidth costs followed over the past two decades, many applications that are currently too expensive to run will become economically practical. That cost compression would unlock enormous new markets.

Third, agentic AI needs to work reliably. The current generation of AI models is impressive at generating text and code, but it is unreliable when asked to take autonomous actions in the real world, managing files, browsing the web, completing multi-step tasks without human supervision. If the next generation of models can perform these tasks reliably, the productivity gains for businesses would be significant enough to produce genuine, measurable ROI rather than the elusive returns most enterprises are experiencing today.

None of these outcomes are impossible. Some of them are probable. But probable in two to four years, at current rates of capital consumption, still requires tens of billions more dollars to be spent before the payoff arrives.

For Africa, the timing of this maturation matters enormously. If reliable, low-cost AI application tools arrive in the next two to three years, Kenya and other leading African tech markets are well-positioned to build on top of them. The mobile money infrastructure, the young, tech-literate population, the established startup ecosystem in Nairobi, and the genuine unmet needs in healthcare, agriculture, and financial services all create conditions for high-impact AI applications. If the technology matures on that timeline, Africa's position as a fast-follower could be an advantage rather than a disadvantage.

If the maturation takes longer, or if the costs remain high enough to exclude markets where average spending power is lower, the continent risks being a consumer of AI tools built elsewhere rather than a creator of tools built for its own context.

My Honest Verdict

The AI industry is not the dot-com bubble. The technology is real, the capabilities are genuine, and the trajectory of improvement is clear. A model that could autonomously find a 17-year-old security vulnerability in FreeBSD, as Anthropic's Claude Mythos did, is not vapor-ware. A coding assistant that measurably reduces the time developers spend on repetitive tasks is not a gimmick. The technology is working.

But the business model for most of the companies building this technology remains unproven at the scale required to justify their valuations. The money is real. The losses are real. The 95% enterprise failure rate for AI pilots is real. The question of what happens when investors run out of patience for losses that are growing, not shrinking, is real.

The most likely resolution is not a single dramatic crash but a slow consolidation. The companies with the strongest enterprise relationships, the most defensible distribution advantages, and the tightest unit economics will survive and eventually thrive. Many of the smaller players, and perhaps some of the larger ones, will not. Nvidia will continue to profit regardless of who wins, because everyone building needs its hardware.

For Kenya and Africa, the pragmatic conclusion is this: do not try to play the game the model companies are playing. You cannot win that game without the compute infrastructure, the capital, and the talent concentration that exists in San Francisco and London. But the application layer is genuinely open. The problems worth solving are specific, the markets are real, and the foundational tools are increasingly accessible. The $10 billion initiative announced in Nairobi, the GPU compute being distributed to African innovators, and the growing digital talent pool all point toward the same opportunity.

The gold rush may or may not make the prospectors rich. But the person who builds the best tool for the conditions on the ground, the specific terrain, the particular mineral, the local knowledge that nobody else has, that person has always had a viable business. That is still true in 2026, and it will still be true when the dust from the current AI frenzy eventually settles.

Everyone is betting trillions. The business case, for most of them, has not arrived yet. But for the builders who are willing to work on problems that are real rather than fashionable, it is closer than the headline numbers suggest.

Related reading: The Myth of Mythos: Why Anthropic Is Selling You Fear Before They Sell You Shares | Google Broke the Web It Built. Now It's Selling You the Pieces. | Your Home Is Now a Data Center