The Largest Investment Cycle Since Railroads: Goldman Sees $1.2 Trillion in Hyperscaler AI Capex for 2027

For months, the central question hanging over the AI boom has been whether the spending is about to crack. Goldman Sachs has an answer, and it is not the one the bears were hoping for: the bank now expects the five largest US hyperscalers — Amazon, Alphabet, Microsoft, Oracle, and Meta — to pour a combined $1.2 trillion into AI infrastructure in 2027, more than half again above the roughly $800 billion projected for 2026 and comfortably ahead of Wall Street’s consensus estimate of $1.1 trillion.

The call, reported by Bloomberg on September 25 and attributed to strategist Ryan Hammond, comes with a historical comparison that is doing a lot of work: relative to GDP, Goldman describes the coming buildout as the largest investment cycle since railroad construction in the 19th century. That framing either signals the scale of a genuine technological transition or the kind of euphoria that precedes one of history’s great capital bonfires — and the honest answer is that nobody yet knows which.

The numbers behind the call

The trajectory Goldman sketches is one of extraordinary levels with decelerating growth. From roughly $150 billion in combined annual hyperscaler spending in 2023, the five companies climbed to an estimated $800 billion in 2026 — a 94 percent jump from 2025. The bank projects $1.2 trillion in 2027 and $1.4 trillion in 2028.

Look at the growth rates rather than the totals, though, and the shape of the curve changes: nearly 100 percent growth in 2026, slowing to 54 percent in 2027 and just 12 percent in 2028. The buildout is not ending; it is maturing. The additional $200 billion expected in 2028 alone is almost as much as these companies spent during all of 2024, and the $1.2 trillion projected for 2027 exceeds their combined spending from 2023 through 2025.

Zoom out further and the stakes get larger. Goldman’s Global Institute and Global Investment Research estimate roughly $7.6 trillion of cumulative AI infrastructure investment from 2026 through 2031, spanning compute, data centers, and power, with a baseline model that reaches $1.6 trillion of annual AI capex by 2031.

The revenue question that will not go away

Here is the tension at the heart of the forecast. To recoup those outlays, Goldman calculates the hyperscalers need to generate roughly $300 billion a year in AI revenue. Current earnings still fall short of that mark. The bullish counterargument is momentum: cloud revenue growth accelerated from 25 percent in 2024 to 48 percent by the second quarter of 2026, and enterprises are demonstrably willing to pay for inference at scale.

Whether revenue growth at AI labs like OpenAI and Anthropic is fast enough to justify the spending remains genuinely unclear. Both companies sit at the center of the expectations — and increasingly the financial instruments — backing this buildout. Goldman also notes that spending now exceeds what the hyperscalers generate from ongoing operations, which points toward more debt financing. JPMorgan, for its part, has estimated that $4.1 trillion in AI-related debt will be issued through 2030. With 10-year Treasury yields hovering near 5.17 percent, up about a full percentage point since January, the cost of that leverage is climbing at exactly the wrong moment for debt-hungry operators.

Physical bottlenecks compound the financing pressure. Goldman flags power, labor, and memory chips as constraints that could slow the buildout further, and the International Energy Agency expects global data-center electricity consumption to roughly double from 485 terawatt-hours in 2025 to 950 TWh in 2030, with AI-focused data centers tripling their draw.

The bottleneck is moving beyond GPUs

Perhaps the most strategically interesting implication of the Goldman analysis is where the next dollar of spending goes. The first wave of the AI boom rewarded companies selling the processors. A GPU sitting in a warehouse, however, produces exactly zero AI revenue — it needs networking, memory, optical connections, cooling, electrical equipment, buildings, and enormous amounts of electricity before it does anything useful.

Constrained supply has already left fingerprints in the margins: Goldman says memory producers’ gross margins have been pushed to roughly 80 percent, more than double their historical average — a signal of extraordinary pricing power, and, implicitly, of the risk that today’s bottleneck profits attract tomorrow’s capacity glut. The bank also cautions that slower capex growth will eventually collide with rising depreciation expenses, potentially diluting AI’s contribution to S&P 500 earnings even if the absolute spending keeps rising.

Context: this fits a pattern of rising estimates

Goldman’s new figures land on a trend line that has bent steadily upward all year. The bank had already flagged in June that consensus estimates were far too low. S&P Global made a structurally similar call in late August — hyperscaler AI capex heading toward $1.3 trillion with cash burn persisting into 2029. Bloomberg Economics has warned the boom could distort national capex figures in smaller economies like Australia. Each new estimate has arrived above the last, and the “consensus” has chased the forecasters rather than the reverse.

That is how bubbles look in their middle chapters, and also how genuine infrastructure transitions look. Railroads, for what it is worth, were both: they transformed economies and ruined a large share of their investors, usually in that order. The difference between the two outcomes was rarely the size of the buildout — it was whether the revenue materialized faster than the financing costs compounded.

What to watch

Three indicators will test the $1.2 trillion thesis over the coming quarters. First, hyperscaler guidance in upcoming earnings — every one of the five has a track record this cycle of raising capex guidance rather than cutting it. Second, AI revenue disclosure: the gap between the ~$300 billion annual revenue need and today’s run rate is the single most important number in the story. Third, the debt markets — with $4.1 trillion of AI-linked issuance projected through 2030 and yields at their highest since 2007, the marginal data center is increasingly financed at the marginal cost of capital.

Goldman’s message, stripped of its qualifiers, is that the AI buildout has outgrown the category of a tech spending cycle and become a macroeconomic event. The five largest spenders are now, collectively, one of the largest capital-allocation decisions in modern history. Whether that reads in retrospect as visionary or infamous depends on a revenue line that is still being written.