American technology companies bet $200 billion on AI whilst profits remain elusive
Sweetgreen's journey from salad seller to 'automation company' reveals broader pattern as Goldman Sachs and Wall Street question massive infrastructure spending without demonstrated returns
In 2021, Sweetgreen spent $70 million acquiring a robotics startup and promptly declared itself an "automation company that serves salads on the side." The salad chain's executives spoke breathlessly about artificial intelligence and technological transformation. Investors, apparently, believed them.
Four years later, Sweetgreen just sold that robotics division for $186 million whilst its actual business collapses. Same-store sales have plummeted for three consecutive quarters—down 9.5% in the most recent, driven by an 11.7% collapse in customer traffic. Net losses have nearly doubled. The stock trades far below its public offering price.
Here's the remarkable bit: Sweetgreen technically made money on the robotics sale. Yet this "success" masks a more interesting failure—one that illuminates a $200 billion question now haunting American technology.
Across the industry, companies have contorted themselves into "AI companies" regardless of business fundamentals, deployed staggering capital before identifying profitable uses, and watched valuations soar on promises rather than results. Now markets are asking awkward questions. The answers aren't encouraging.
The scale of the gamble
Microsoft will spend roughly $80 billion on AI infrastructure this fiscal year. That exceeds Luxembourg's entire GDP. Last quarter alone: $22.6 billion, with chief financial officer Amy Hood confirming "nearly all" went to AI.
Amazon, Meta, and Google represent hundreds of billions more. Zuckerberg announced plans for an AI data centre that "could cover a significant part of Manhattan." Amazon's cloud division spent $26.3 billion in Q4 2024, "the vast majority" on AI, according to chief executive Andy Jassy.
Nvidia, selling the graphics processors that power these systems, briefly topped $3 trillion in market value. The entire technology ecosystem has oriented itself around one assumption: AI will generate returns justifying this spending.
Yet the sector's most emblematic company reveals economics that don't work. OpenAI, creator of ChatGPT, lost roughly $5 billion in 2024 on $3.7 billion revenue. Recent disclosures suggest quarterly losses have accelerated to $11.5 billion. The company projects burning through $115 billion by 2029.
OpenAI claims 800 million weekly ChatGPT users—adoption at extraordinary speed and scale. But the fundamentals remain broken. Training models costs approximately $3 billion. Running them costs another $2 billion. These figures grow with usage. Counterintuitively, every additional user increases the cash burn.
Spending without returns
BCG research from late 2024 found that just 4% of companies have achieved "cutting-edge" AI capabilities enterprise-wide. The other 74% have yet to show tangible value despite widespread investment.
Two-thirds of businesses remain "stuck in AI pilot mode," according to an Informatica survey of 600 data leaders. More damning: 97% struggle to demonstrate business value from generative AI initiatives.
This extends beyond startups. Microsoft, despite its massive spending, recently experienced its worst stock performance in over a decade—shares down 8.6% over eight days, erasing $350 billion in market value. Azure continues growing, but analysts increasingly question whether AI spending will generate commensurate returns.
The measurement challenge itself reveals the problem. Unable to demonstrate hard financial returns, companies pivot to "soft ROI"—qualitative productivity improvements, employee satisfaction, potential future benefits. Deloitte reports 78% of companies plan increased AI spending. Most cannot articulate clear success metrics beyond vague productivity gains.
When Goldman Sachs tested AI for updating historical data in company models, the technology worked. At six times the cost of manual methods.
MIT research documented that 95% of AI pilot projects fail to yield meaningful results despite more than $40 billion in generative AI investment. The pattern holds: enormous spending, elusive returns.
Wall Street grows sceptical
Jim Covello doesn't mince words. Goldman Sachs' head of global equity research stated his assessment bluntly: "AI technology is exceptionally expensive, and to justify those costs, the technology must be able to solve complex problems, which it isn't designed to do."
Then Covello twisted the knife: "People generally substantially overestimate what the technology is capable of today. Even basic summarisation tasks often yield illegible and nonsensical results. Despite its expensive price tag, the technology is nowhere near where it needs to be for even such basic tasks."
This from a man whose job involves evaluating technology investments professionally. His conclusion challenges the core assumption underlying AI valuations.
Goldman Sachs Research titled their June 2024 paper unambiguously: "Gen AI: too much spend, too little benefit?" The analysis noted tech giants plan roughly $1 trillion on AI infrastructure in coming years, yet "this spending has little to show for it so far."
Daron Acemoglu, Institute Professor at MIT, projects generative AI will have muted impacts on labour productivity and GDP growth over the next decade. His forecast directly contradicts the transformational claims driving current valuations.
The International Monetary Fund weighed in last October. Chief economist Pierre-Olivier Gourinchas drew explicit parallels: "There are many similarities between the late 1990s internet stock bubble and the current AI boom, with both eras pushing stock valuations and capital gains wealth to new heights." The promise of transformative technology, he noted, may not meet market expectations in the near term—triggering valuation crashes.
Gourinchas emphasised one crucial distinction: AI investment has increased just 0.4% of US GDP since 2022, compared to the dot-com era's 1.2% increase between 1995 and 2000. The current boom, whilst substantial, remains smaller macroeconomically. Today's spending comes from cash-rich technology companies, not debt-fuelled startups.
Still. The parallel haunts every analysis.
Echoes of excess
The late 1990s telecommunications companies laid vast fibre-optic networks based on internet traffic projections. The infrastructure proved decades ahead of demand. WorldCom, Global Crossing, and Qwest collapsed. Corning's stock crashed from nearly $100 to roughly $1 in two years. Ciena's revenue plummeted from $1.6 billion to $300 million almost overnight.
Yet that prematurely built fibre eventually proved valuable as internet usage expanded. Today's corporate giants—Microsoft, Google, Meta, Amazon—possess resources their dot-com predecessors lacked. They generate enormous cash flows that can sustain investment through extended periods without positive returns.
The question remains whether current AI technology has reached a capability plateau insufficient to justify the infrastructure. OpenAI's mounting losses despite 800 million weekly users suggest the economics don't function at realistic scale. The assumption that deploying more computing power automatically generates proportionately more capability may prove flawed.
The prisoner's dilemma
Companies find themselves trapped. Investment continues not because executives see clear returns, but because competitors are investing. Goldman Sachs documented this explicitly: stock prices already reflect assumed productivity gains "before it materializes, raising the risk of overpaying."
No individual company can afford to stop. Microsoft cannot cede infrastructure leadership to Amazon. Meta cannot allow Google's AI capabilities to advance unchallenged. The industry has locked itself into competitive spending regardless of profitability.
The market correction from early November 2025 suggests investors are pricing this in. Roughly $1 trillion in market value evaporated from major technology stocks over one week. Nvidia, Meta, Oracle, and Palantir all declined significantly. Microsoft's slump represented its worst performance since 2011.
These movements don't necessarily portend catastrophe. Markets regularly reassess valuations when evidence contradicts narratives. The dot-com bubble burst partly because many companies had no revenue. Today's technology giants generate substantial income—Microsoft's Azure hit an $86 billion annual run rate. These are real businesses with genuine customers, not pure speculation.
What markets question is the marginal return on the next $80 billion of AI spending. Will Microsoft's enormous investment generate proportionate Azure revenue increases? Can Meta monetise its AI capabilities sufficiently? Does OpenAI possess a path to profitability, or will it require perpetual subsidies?
The answers remain uncertain. Goldman Sachs analysts who question AI returns acknowledge they could be wrong—that breakthrough applications might emerge within 18 months justifying current spending. History shows transformative technologies often take longer than enthusiasts project, yet ultimately prove valuable.
What seems clear: the phase of uncritical enthusiasm has ended. Companies can no longer simply announce AI initiatives and watch stock prices rise. Sweetgreen's trajectory—from "automation company that serves salads" back to struggling restaurant chain divesting its technology—offers a cautionary tale about gaps between technological positioning and business fundamentals.
American technology companies have placed an extraordinary bet on artificial intelligence. Whether it pays off may determine not just individual fortunes, but productivity growth and economic development trajectories for the coming decade. The capital deployed ensures the outcome matters enormously. The mounting scepticism from analysts, economists, and investors suggests the outcome remains far from certain.