$650 billion buys you hope

Six hundred and fifty billion dollars. That's how much Alphabet, Amazon, Meta, and Microsoft are expected to collectively spend on AI infrastructure in 2026, according to Bridgewater Associates. Up from roughly $410 billion in 2025, a 58% increase in a single year. To put that in perspective, the Apollo program cost about $257 billion in today's dollars. The entire GDP of Sweden is around $600 billion. Four companies are now outspending a mid-sized European nation, annually, on data centers and chips. The question is no longer whether Big Tech believes in AI. It's whether belief alone can justify the bill.

The numbers behind the bet

The breakdown is staggering. Amazon leads with a $200 billion capex plan, nearly all of it earmarked for AWS to handle surging AI workloads. Alphabet is forecasting $175 to $185 billion, doubling down on Gemini models, Vertex AI, and Google Cloud expansion. Meta projects $115 to $135 billion, focused on Llama models and AI-powered ad infrastructure. Microsoft rounds it out at roughly $145 billion, channeled into Azure and its OpenAI partnership. Together, these companies generate more than $1.5 trillion in annual revenue. So $650 billion represents roughly 15 to 20 percent of their combined top line. That's not a side bet. That's a restructuring of where the money goes.

The cash flow problem

Here's where things get uncomfortable. In 2024, these four companies generated a combined $237 billion in free cash flow. By 2025, that dropped to $200 billion. The projections for 2026 are far worse. Pivotal Research estimates Alphabet's free cash flow will plummet nearly 90%, from $73.3 billion to just $8.2 billion. Morgan Stanley projects Amazon will swing to negative free cash flow of $17 billion, with Bank of America estimating the deficit could reach $28 billion. Barclays sees a similar 90% drop for Meta, and forecasts the company will post negative free cash flow in 2027 and 2028. Amazon has already filed with the SEC indicating it may need to raise equity and debt to fund its build-out. Alphabet held a $25 billion bond sale in late 2025, quadrupling its long-term debt to $46.5 billion. These are not the balance sheet moves of companies swimming in excess cash. These are the moves of companies mortgaging the present to own the future.

Bridgewater's warning

Bridgewater co-chief investment officer Greg Jensen didn't mince words in a letter to clients. The AI boom, he wrote, has entered a "more dangerous phase," marked by exponentially rising investments in physical infrastructure and growing reliance on outside capital. The pattern is familiar. Compute demand continues to outpace supply, driving hyperscalers to invest faster and faster to try to get ahead of the curve. But Jensen warned that this scale of spending poses real risks if market conditions shift. AI startups like Anthropic and OpenAI may need significant breakthroughs to justify their future funding and valuations, and any stumble could ripple outward. This is the dot-com playbook: build the infrastructure first, figure out monetization later. The difference is that the infrastructure costs are orders of magnitude higher this time around.

The counter-argument: rational fear

There's a strong case that this spending is entirely rational, just not in the way traditional financial analysis would frame it. Microsoft reported Azure cloud services growing 33% year-over-year in Q3 FY25, with AI contributing 16 percentage points of that growth. Google Cloud revenue grew 48% year-over-year to $17.7 billion in Q4 2025. Amazon Web Services posted its fastest growth in 13 quarters. The demand signal is real. The logic goes something like this: whoever owns the compute layer owns the next decade. Every business building AI agents, every developer "vibe coding" applications from text prompts, every enterprise deploying AI-powered workflows, they all need compute. The hyperscalers are betting that AI infrastructure will be as foundational as the internet itself, and that the companies supplying it will extract value from every layer above. As one analyst put it, the bigger risk isn't overconfidence. It's under-investing and leaving core products like search, cloud, and advertising exposed to disruption.

But compute isn't the whole story

There's a gap in this narrative that doesn't get enough attention. Raw compute capacity is necessary, but it's not sufficient. You can have all the GPUs in the world, but if you can't distribute the intelligence effectively, it's wasted capex. The real moat isn't in the data center. It's in the ability to turn compute into products that people actually use, in distribution. The companies that will win aren't necessarily the ones spending the most on infrastructure. They're the ones who can get AI into the hands of users in ways that create sticky, recurring value. Meta understands this intuitively, which is why its AI ad infrastructure has helped drive the company past $200 billion in annual revenue. Microsoft gets it too, embedding Copilot across its entire productivity suite. The question for each of these companies is whether their distribution advantage grows proportionally with their infrastructure spend, or whether they're building capacity that sits idle.

What this means for everyone else

If you're not one of the four companies writing these checks, the implications are sobering. The EU has unveiled a €200 billion AI action plan. Japan has allocated ¥1 trillion annually for AI and semiconductor development. South Korea's 2026 AI budget stands at 9.9 trillion won. And then there's Singapore. The government recently committed S$1 billion over five years for AI public research, alongside S$150 million for its Enterprise Compute Initiative to help companies build AI capabilities. There's also a sovereign AI initiative with $740 million allocated for national AI capabilities in healthcare, urban planning, and defense. These are meaningful investments for a small nation. But they're rounding errors compared to what a single hyperscaler spends in a quarter. Singapore's strategy, wisely, isn't about competing on raw compute. It's about becoming the regulatory and governance hub for AI in Asia Pacific, the place where companies design, test, and deploy AI systems across the region. When you can't outspend, you outmaneuver.

The price of not wanting to be left behind

There's a particular kind of spending that happens when the downside of being wrong feels existential. It's not traditional capital allocation. It's closer to an arms race, where the cost of falling behind exceeds the cost of overspending. That's what $650 billion buys you. Not certainty. Not guaranteed returns. Hope. The hope that AI will be as transformative as its proponents claim. The hope that the revenue will eventually catch up to the capex. The hope that building now, at any cost, is better than building later and finding out someone else already owns the future. History suggests that some of this spending will be wasted. The dot-com era built fiber optic networks that sat dark for years before demand caught up. But those networks eventually became the backbone of the modern internet. The infrastructure wasn't wasted, it was just early. The question for 2026 isn't whether $650 billion is too much. It's whether the companies writing the checks can survive the gap between investment and return. Because the "dangerous phase" Bridgewater describes isn't about the technology failing. It's about the financing failing before the technology has time to prove itself. And that's a bet no amount of compute can hedge.

References