Goldman Says $7.6 Trillion Is Coming. Micron Already Proved It.
Three stories landed this week that, taken together, draw the clearest picture yet of where AI capital is headed — and where the cracks are forming.
Goldman Sachs Just Put a Number on the AI Buildout: $7.6 Trillion
Goldman Sachs updated its AI infrastructure outlook on September 25, now expecting hyperscaler capital expenditures to hit $1.2 trillion in 2027 and $1.4 trillion by 2028, up from roughly $800 billion this year. That builds on the bank’s June projection of $7.6 trillion in cumulative AI infrastructure investment from 2026 through 2031.
But the headline number obscures the tension underneath. Goldman simultaneously quantified the gap between what’s being spent and what’s being earned: the five largest hyperscalers — Amazon, Alphabet, Microsoft, Meta, and Oracle — are collectively spending $800 billion on AI infrastructure in 2026 while generating only $70 billion above their pre-AI revenue baseline. They need approximately $300 billion in annual AI revenue just to break even, leaving a structural shortfall of $230 billion per year.
The investment thesis is shifting. Goldman’s note argues that the first AI boom rewarded chip companies, but the next stage will reward the businesses supplying everything required to keep those chips running — memory, networking, optics, and power infrastructure. Memory producers’ gross margins have already expanded to roughly 80%, more than double their historical average. Global data-center electricity consumption is projected to double from 485 to 950 terawatt-hours by 2030.
For investors, this is both a roadmap and a warning. The spending is real and accelerating, but the revenue to justify it is not yet there. The companies that control the bottlenecks — not just the compute — are where the structural advantage is building.
Micron Reports Wednesday — and the Bar Is the Biggest Quarter in Memory History
Micron Technology (MU) reports fiscal Q4 results on Wednesday, September 30, and the setup matters. Its last print, fiscal Q3 in late June, validated every thesis about AI-driven memory demand. Revenue hit $41.46 billion — up 346% year-over-year — with non-GAAP earnings of $25.11 per share, crushing consensus by $4.40. Gross margins reached a company record of 84.9%, more than doubling from 39% a year prior.
The numbers alone are remarkable, but the forward visibility is what matters most. CEO Sanjay Mehrotra announced 16 Strategic Customer Agreements covering approximately $100 billion in minimum-price revenue commitments with cash deposits of $22 billion. These are take-or-pay contracts that lock in demand through 2030. HBM4, Micron’s next-generation high-bandwidth memory, is ramping twice as fast as HBM3E, with both product lines fully booked through 2027.
Q4 guidance came in at $50 billion in revenue with gross margins of approximately 86% — both well above Wall Street expectations at the time. That is the number Micron has to hit on Wednesday.
Micron is the clearest proof point for Goldman’s infrastructure thesis. When memory producers are printing 85% gross margins and locking in $100 billion in guaranteed revenue, the AI spending cycle is not theoretical — it is being contractually committed at unprecedented scale.
The People Building Frontier AI Are Asking Regulators to Slow Down
On September 28, more than 20 researchers from Anthropic, OpenAI, Meta, and Microsoft — including Nobel laureate Geoffrey Hinton and Anthropic co-founder Jack Clark — published a paper calling for international oversight of self-improving AI systems.
The paper describes recursive self-improvement as the point where an AI system can enhance its own capabilities, potentially initiating a feedback loop that accelerates advancement beyond human oversight capacity. The authors call for mandatory independent evaluations before deploying frontier models, global technical standards for safety testing, and international coordination frameworks to prevent regulatory arbitrage.
This is not academic hand-wringing. It comes weeks before Anthropic’s expected IPO in mid-October, with a valuation target of up to $2 trillion. The timing matters: the company that is about to become the largest AI IPO in history is simultaneously telling regulators that the technology it sells may soon need guardrails its creators cannot provide alone.
For investors, the regulatory signal is structural. Companies that invest in safety infrastructure and build compliance into their development process will have a moat if regulation arrives. Those that treat safety as an afterthought face the kind of sudden regulatory risk that has historically compressed valuations overnight.
What It Means for Investors
The $7.6 trillion spending forecast, Micron’s record quarter, and the safety call from frontier AI leaders all point to the same conclusion: the AI buildout is entering its infrastructure phase, and the rules governing it are about to change.
The capital is flowing — that much is settled. Goldman’s numbers and Micron’s contracts prove it — and Wednesday’s report is the next test. But the sustainability of that spending depends on revenue catching up to capex, and the companies best positioned are those controlling physical bottlenecks: memory, networking, power, and cooling.
Meanwhile, the regulatory landscape is shifting from permissive to proactive. The architects of frontier AI are no longer waiting for governments to act — they are actively shaping the framework. For long-term investors, this means the competitive moat in AI is not just about model performance anymore. It includes safety infrastructure, regulatory positioning, and the ability to operate under rules that do not yet exist.
The AI trade is not over. But it is no longer just about who builds the fastest model. It is about who builds the infrastructure to sustain it — and who earns the right to keep building.
MasicotAI — Tracking the intersection of artificial intelligence and economic reality.