#GoldmanSees1.2TAICapex
About GoldmanSees1.2TAICapex
Goldman Sachs expects combined 2027 capital spending by Meta, Microsoft, Alphabet, Amazon and Oracle to reach about $1.2 trillion, up from roughly $800 billion in 2026, largely driven by AI infrastructure. The buildout may support demand for chips, memory, data centers, power and cloud services, but monetization remains the key test. Can AI applications generate enough revenue and cash flow to justify the rising investment?
Hot
Latest
GoldmanSees1.2TAICapex Popular posts
The chart takes AI capex (Amazon, Google, Meta, Microsoft, Oracle + SpaceX, accumulated through each year) as an input, then converts it into the perpetual annual revenue needed for a 10% return under each model’s assumptions about depreciation, opex, and return metric. Annual revenue required is on the y-axis; will exceed $1 trillion per year in 2027.
The “Is AI a bubble?” narrative has cooled off a bit in 2026—partly bc of AI revenue growth in the age of agents, and partly because the discourse puck has moved to AI safety.
But the emerging consensus that new AI revenue earned by the hyperscalers will have to approach $3 trillion annually by the end of the decade is still unbelievably daunting.
Blockspace
Goldman Sachs: latest hyperscaler capex estimates show $996 billion in 2027
Neoclouds and former Bitcoin miners have secured meaningful hyperscaler capex since 2025
$NBIS: Meta + Microsoft
$CRWV: Google, Meta, Microsoft
$CIFR: AWS + Fluidstack (Google-backed)
$IREN: Microsoft
$WULF: Fluidstack (Google-backed)
$HUT: Fluidstack (Google-backed)
$10.3 trillion. That's Brookings' estimate for US AI infrastructure investment from 2025 to 2032, highlighted by @Axios.
People talk a lot about whether AI is a bubble. This Brookings paper asks a different, though related, question: could financing the buildout create systemic risk? An AI bubble and a systemic risk event aren't the same thing. A bubble is when investment or asset prices run ahead of the returns. Systemic risk is when losses spread beyond the projects that made the bad bets and disrupt the wider financial system.
So what would need to happen? The paper points to concentrated exposure across banks, insurers and funds, all relying on the same AI tenants. If weaker AI demand hit utilization, data-centre and GPU values, and refinancing at once, losses could spread from projects to lenders and investors.
The author says it's premature to conclude we're there - and I agree. But much of this financing sits in structures that make it hard to see who ultimately holds the risk... Regardless, we're still just warming up and its not going to slow down anytime soon.
TL/DR: Goldman thinks $1.42 trillion in revenue over 2028-30, or $11.6BN per GW - is clearable (required for 15% ROIC), with 59% of it already sitting in cloud backlogs and management teams claiming paybacks of anywhere from one to three years
And here is the counter: the hurdle is clearable if compute stays scarce, if GPU and token prices don't deflate, if a 5-year useful life is real, if a handful of AI labs keep paying their bills, and if the $3 trillion off-balance sheet iceberg doesn't need a return of its own.
i highlighted Meta last week and the thesis is starting to get more interesting
their AI spend is beginning to show up across the actual business, especially ads, while the consumer side keeps expanding with products like Meta AI
the market has spent a lot of time focused on who owns the best model
distribution, monetization and the ability to ship AI directly to billions of users matter just as much
Meta is one of the names i’m watching closely here
Ansem 🐂🀄️
from my perspective, Meta is still pretty under-discussed in the AI race
> open-source models keep getting more efficient
> distribution becomes a huge edge as models improve
> Meta already has billions of users across its products
> better AI can improve personalization + ad targeting
its open-source AI strategy gives it a different position from other big labs, one of the larger AI companies i’m paying more attention to right now
🚨 AI IS GETTING CHEAPER AT A RIDICULOUS SPEED
ACCORDING TO EPOCH AI, THE SAME LEVEL OF AI PERFORMANCE HAS BECOME ~13× CHEAPER EACH YEAR SINCE 2023
FASTER THAN ELECTRICITY, COMPUTE, BATTERIES OR DNA SEQUENCING ON THIS CHART
BUT RUNNING AI STILL TAKES CHIPS, POWER AND DATA CENTERS
COMPANIES ARE SPENDING BILLIONS TO SELL SOMETHING THAT KEEPS GETTING CHEAPER
IT’S STARTING TO LOOK LIKE A VERY EXPENSIVE CHARITY
EXCLUSIVE: Subconscious is tackling one of the biggest challenges facing long-running AI agents: inference cost.
@subsysdev has raised $5.1 million to build an inference platform designed specifically for agents, using dynamic context compression and caching to make long-running workloads faster and cheaper. The company says its technology can reduce inference costs by up to 80%, while extending effective context windows beyond 5 million tokens. For businesses scaling agentic AI, that could make it more practical to run agents for longer periods on increasingly complex tasks.
@deantak spoke to @thejackobrien, CEO of Subconscious, about the company’s technology, its focus on long-running agents and the economics of powering the next generation of AI applications.
“Most inference companies build one-size-fits-all systems for chats, one-shot requests, and agents alike. But agents are much harder to serve well. We built our whole stack around long-running agents, and our core technology comes from novel MIT research, which gives us a year-plus head start,” O’Brien said.
Read the full story on @GamesBeat:
#AI #AgenticAI #Technology
