AI’s Reckoning

The grand promise of Artificial General Intelligence is collapsing, pushing AI into a “Trough of Disillusionment.” Trillions in investments now face a reckoning as the AI bubble threatens to burst, with potential economic contagion.

The air was thick with anticipation, a modern-day Los Alamos moment for the digital age.

Tech titans and eager investors, their collective breath held, watched as OpenAI CEO Sam Altman prepared to unveil GPT-5, the latest iteration of what many believed was humanity’s express train to Artificial General Intelligence.

This wasn’t just another product launch; it was supposed to be a leap towards a machine smarter than all of us, a godlike intelligence heralding a new epoch.

The promise, relentlessly hyped for years, was of the Singularity, a gentle, yet apocalyptic, transformation already underway.

But the grand reveal was, to put it mildly, a dud.

Instead of a quantum jump towards superintelligence, what the world got was a “model router.”

OpenAI pitched it as a clever tool for efficiency, dynamically selecting computational power for queries, translating to faster, cheaper answers.

Yet, as writer Ed Zitron astutely observed, this offered little real value to the customer while conveniently lowering OpenAI’s cloud bill.

It felt like a cruel cosmic joke: we were promised AGI, and all we received was a glorified traffic cop for algorithms.

The online community, ever quick with a barb, christened the day “Gary Marcus Day,” a nod to the eminent AI scientist who has, for decades, warned of the structural limitations and inherent fallibility of this very dream.

Weeks later, Altman himself conceded the rollout was poorly done, offering the familiar refrain that the next model, GPT-6, would surely deliver.

He spoke of spending trillions on data centers, even as he, with a straight face, cautioned “overexcited” investors – but certainly not himself – about an impending AI bubble, conjuring specters of the dot-com crash.

The irony was palpable.

If the AGI dream is merely delayed, let alone dead, the investor nightmare is only just beginning.

This isn’t just about Silicon Valley’s latest hubris.

The U.S. economy, dangerously intertwined with Big Tech, has priced its investments on the very promise of AGI.

The question isn’t just what happens to the Valley, but to the global economy, if this promised superintelligence never arrives.

A meme, “What did Ilya see?”, circulated online after Ilya Sutskever, a co-founder of OpenAI and the lead researcher behind ChatGPT, abruptly departed.

Sutskever, the man arguably most responsible for the generative AI revolution, had signed on to the board’s attempt to oust Altman for a lack of candor.

People wondered: why would a pioneer abandon his company at the precipice of humanity’s “final invention”?

Perhaps, as the whispers suggested, Ilya saw something far more troubling than merely Sam’s trustworthiness.

Soon after his exit, Sutskever delivered a talk that, while technical, was nothing short of heresy within the industry.

He effectively declared the end of pre-training scaling – the almost magical ability of large language models (LLMs) to get smarter, more capable, and even learn new skills simply by getting bigger.

The leap from GPT-2 to GPT-3, which birthed ChatGPT, was precisely this strategy.

GPT-4 followed suit.

It seemed so simple: more data, better models.

But judging by GPT-5’s underwhelming performance, that era of effortless progress, as Ilya seemingly foresaw, is drawing to a close.

Dr. Marcus, enjoying what one could only imagine as a well-deserved victory lap in his convertible, confirmed this.

The GPT-5 launch, he asserted, proved that LLMs are not the “royal road to AGI.” He has long critiqued scaling AI, pointing to LLMs’ lack of symbolic reasoning and their notorious “hallucinations.”

Unlike traditional software, LLMs are probabilistic, meaning the link between their training data and their output is inherently unstable.

This instability is both their power and their fatal flaw.

It explains why ChatGPT can assist with Ph.D.-level research one day and confidently miscount the R’s in “strawberry” or struggle with basic numerical comparisons the next.

Hallucinations are not bugs to be patched; they are a feature of this fundamental structural flaw, unresolvable by merely making models bigger.

To be clear, LLMs are a powerful, useful technology.

Their ability to understand natural language, generate creative content, and learn from mistakes are genuine achievements.

They offer real productivity gains, even if not the apocalyptic transformation once envisioned.

Marcus views them as “one part of AGI,” but crucially, they lack symbolic reasoning – the ability to definitively code things as “true” or “false.”

He has long advocated for neurosymbolic approaches, a potential path forward now that the scaling laws are faltering.

But such new paradigms are years away from replicating the breathtaking progress of the recent past.

We are, as Marcus put it, “in an era of diminishing returns for AI.”

While techniques like “test-time scaling” offer incremental improvements, they cannot match the seismic leaps of pre-training.

AGI, by any definition, remains elusive with current methods. Consulting giant Gartner has already moved generative AI off its “Peak of Inflated Expectations” and into the “Trough of Disillusionment,” a phase often marked by decreased investment and interest.

If the rapid progress from scaling laws has indeed ended, the implications are staggering.

The world-historical bets investors have placed on AI will begin to sour.

Silicon Valley, despite its self-proclaimed contrarian spirit, has stampeded off a cliff, pouring a record $110 billion into AI startups last year, representing an astonishing 42 percent of all VC funding.

This has fueled 498 AI unicorns, translating $350 billion in VC investments since 2021 into a staggering $2.7 trillion valuation.

These valuations are built on thin air if the pace of progress cannot be sustained.

Consider OpenAI itself.

Despite projections of $20 billion in annual recurring revenue this year, it expects to lose $5 billion.

By 2026, losses are projected to hit $14 billion annually, even with rising revenue.

OpenAI doesn’t foresee breaking even until it clears $100 billion in revenue by 2029 – a third of what giants like Microsoft, Amazon, and Google make annually.

Zitron labels this the “rot economy”: companies that “burn billions to lose billions.”

It’s difficult to justify a company planning to lose over $100 billion this decade being valued like Coca-Cola or twice that of Spotify.

One desperate pivot for cash is OpenAI’s proposed $500 billion “Stargate” supercluster, hyped as a new Manhattan Project for AI infrastructure.

The comparison is apt, perhaps too apt.

If AI data centers are the new railroads or fiber-optic cables, history offers a grim prognosis.

The largest U.S. railroad went bankrupt.

Two of the three largest fiber-optic cable builders collapsed during the dot-com bust.

Current AI infrastructure spending, projected to race past $300 billion this year and reach an estimated $1 trillion annually in capex by 2028, already exceeds the dot-com era’s telecom bubble peak.

Each day without AGI brings us closer to a financial reckoning.

The valuations of companies like OpenAI and Nvidia hinge on this projected infrastructure being fully built and utilized.

No one spends a trillion dollars a year on capex without AGI-sized returns.

The data center buildout increasingly feels like a literal and figurative waste of energy, and investors may simply walk away.

As the saying goes about the dot-com era, “The internet hasn’t disappeared, but its investors did.”

While data centers might be the first to feel the pop, other key players are equally vulnerable.

Nvidia, the “picks and shovels” provider of the AI gold rush, akin to Cisco during the dot-com era, has seen its stock soar by 3,653 percent in six years, briefly making it the world’s most valuable company.

But if the limitless demand for its GPUs falters, the entire company could stagger.

The NASDAQ lost 80 percent of its value during the dot-com crash, and so did Cisco.

How would the market handle a similar decline in Nvidia and a precipitous drop in Big Tech spending?

Then there’s Tesla, Wall Street’s ur-AGI bet.

Investors have swallowed Elon Musk’s decade-long promises of fully self-driving cars, humanoid robots, and Martian colonies, granting Tesla a price-to-earnings multiple of 188, compared to BYD’s 27.

This speculative valuation persists even as Tesla’s core business shows signs of decline, with falling sales, the much-maligned Cybertruck, and the loss of crucial tax and carbon credits.

Without the mythical robots or robotaxis, the idea of Tesla being worth more than every other automaker combined becomes increasingly suspect.

The numbers for the broader U.S. economy are truly unsettling.

The “Magnificent Seven” tech giants now account for 34 percent of the S&P 500.

In the first half of 2025, Big Tech’s AI infrastructure spending contributed more to U.S. GDP growth than consumer spending, a flashing red light for an economy where consumers typically drive 70 percent of activity.

The economic tide seems to be receding everywhere except the tech economy, which itself might be a mirage.

Capital can be patient, but the AGI bet feels like another iteration of the “rot economy” playbook that gave us WeWork, NFTs, blockchain, the metaverse, and Theranos.

Yet, Wall Street, pension funds, private capital, and sovereign wealth funds all bought in, seduced by the siren song of transformation.

In 2007, a collapse in the subprime mortgage market triggered the Great Recession, an international conflagration.

If the AGI bubble bursts, Silicon Valley and the stock market face billions – potentially trillions – in dead capital, amidst their most fragile state in years.

Unlike during COVID, no same-month bailouts can be expected.

When the AGI bubble bursts, can the contagion be contained?

The question hangs heavy, a chilling echo in the digital air.