The Wall Around AI Infrastructure

00:00:04 Alphabet is asking Wall Street for $85 billion. Eighty-five billion — in fresh capital. They've already secured more than $55 billion in debt since November. This latest raise splits that into $80 billion in equity sales and a $5 billion piece of additional funding.

00:00:23 There's a Berkshire Hathaway $10 billion investment in the mix, which is notable but also familiar at this point — Warren Buffett has been buying Alphabet stock heavily. The company's free cash flow will turn negative for the next few years as AI capex ramps, according to Melius Research.

00:00:43 The stock has dropped for four straight weeks going into this announcement. That might sound minor for a company that's up 120% over the past year. But Dan Niles of Niles Investment Management put it bluntly: "I never thought Google would need to hit the public markets to raise money to fund their spending."

00:01:09 It's that Alphabet has been Wall Street's favorite megacap tech name over the past year — and now they're raising equity specifically because the market isn't giving them enough leverage on debt alone. CFO Anat Ashkenazi called it a "strategic proactive move." The pitch is that massive access to global capital markets is becoming a competitive advantage given the scale of AI demand.

00:01:37 Sundar Pichai said enterprise demand is "meaningfully exceeding" Alphabet's available supply. But free cash flow turning negative while capex doubles from last year is what happens when infrastructure spending outpaces the business models it's supposed to support.

00:01:56 Google Cloud revenue did jump 63% year-over-year to $20 billion, and AI solutions are now the largest contributor to cloud growth for the first time. That's real. But at what point does the capex stop being a signal of demand and start being a signal that every company in this space is running faster just to stand still?

00:02:19 New York just passed a one-year moratorium on new large-scale data center permits. It's the first state in the country to do this. Governor Kathy Hochul hasn't said whether she'll sign it, but Assembly Speaker Carl Heastie told reporters, "We intend to pass it." The bill bundles several measures: a permit freeze, an environmental impact review requirement from the Department of Environmental Conservation covering water use and electricity consumption, and — this one matters — creation of a separate utility rate class for data centers.

00:02:59 Right now, data centers in many states pay the same commercial utility rates as everyone else. The cost of upgrading electrical infrastructure to serve them gets spread across all customer bills. The new bill tries to isolate that cost. The math behind this is brutal and specific.

00:03:20 A Bloomberg analysis cited in the legislation found that 70 percent of places with rising electric rates are within 50 miles of significant data center activity. In high-concentration areas, electricity prices jumped 267 percent over the past five years. The International Energy Agency reported that global electricity demand from data centers soared 17 percent in 2025 alone.

00:03:48 U.S. data centers' total combined electricity demand is projected to nearly double between 2025 and 2028, rising from 80 to 150 gigawatts. That's roughly equivalent to adding a country with Spain's energy needs in just three years. Consumer Reports surveyed 2,146 adults in November 2025 and found that 78 percent of Americans were somewhat or very concerned about data centers driving up their energy bills.

00:04:19 This isn't abstract anymore. People are seeing it on their electric bills. The compressed timeline here is worth noting: lawmakers consolidated competing proposals into a one-year measure at the end of the legislative session because Democrats acknowledged they needed to show voters action before the November election.

00:04:43 That's political reality layered on top of infrastructure stress.

00:04:47 While New York pauses, someone else is pouring concrete elsewhere. AirTrunk — the Australian data center operator backed by Blackstone — announced it will invest $30 billion in India by 2030, building 5 gigawatts of new capacity. They already entered India earlier this year through the acquisition of Lumina CloudInfra and have a pipeline of about 600 megawatts across Mumbai, Chennai, and Hyderabad.

00:05:17 The Raigad Pen Growth Center project alone is planned for 3 gigawatts with roughly ₹2 trillion in investment. India's data center capacity is projected to rise from about 1.5 gigawatts today to as much as 8 gigawatts by 2030, according to Bernstein research. The Indian government offered foreign cloud providers tax exemptions through 2047 on services sold overseas if those workloads run from Indian data centers.

00:05:47 This connects to a piece in Rest of World about Silicon Valley's lure fading for India's tech talent — and the infrastructure story is just the capital side of that same shift. India startups pay 50 to 75 percent of what Microsoft, Google, or Meta pay for equivalent roles.

00:06:08 But as an early-stage AI company founder put it: getting a foot in the door sooner comes with stock options and performance incentives, and "a lot of people are willing to take that sort of 50 percent sacrifice today for the greater pot at the end." A survey of over 1,200 Indian tech workers found that employees at Microsoft, Amazon, Oracle, and Google are willing to work for homegrown firms if laid off.

00:06:52 India has "only one qualified engineer available for every 10 open generative AI positions," a staffing firm CEO told Rest of World. When supply is this constrained, every credible employer taps the same shallow pool. The talent competition in the five-to-15-year experience bracket — AI, machine learning, cloud engineering, platform architecture — is particularly hot.

00:07:19 The geographic shift in AI infrastructure isn't just about labor costs anymore. It's tax incentives, energy availability, and where the engineers who will actually build and run these systems decide to live.

00:07:34 Two items today circle the same question from different directions. Paul Graham tweeted: "If big companies can't make a net return on their LLM token costs, that doesn't mean it's impossible to. In fact this is exactly what you'd expect to happen with a new technology.

00:07:53 Incumbents can't use it well, and are replaced by upstarts who can." It's the kind of tweet that reads as obvious in retrospect — but also as someone who watched this pattern play out before, across different technology waves. The underlying reality is stark: even with models getting better and inference costs dropping, the revenue being generated by large language model-powered features hasn't caught up with the cost of running them.

00:08:26 Alphabet itself acknowledged this indirectly when Google Cloud's CFO said AI solutions are now "the largest contributor to cloud growth" but also flagged that Gemini serving costs were reduced 78 percent since last year — which is a cost-reduction narrative, not a revenue-growth one.

00:08:46 The second item lands on the same hardware question. A research team including Huawei says it used Ascend 910C chips for DeepSeek V4 Pro's post-training phase, despite U.S. sanctions. Elon Musk responded: "They need some indigenous GPUs 😂" — 3,000 likes. The Forbes piece on Anthropic, OpenAI, and Nvidia frames this the other way around.

00:09:10 Dario Amodei and Jensen Huang had a public clash over China chip policy, with each calling the other's position "crazy" or "stupid." The article notes that export rules have changed where Nvidia can sell its most advanced chips but also triggered a rush of domestic buying in countries hit by restrictions.

00:09:31 Anthropic secured more than 220,000 Nvidia GPUs through a SpaceX deal. That's the dependency chain spelled out plainly: model companies can't exist without chip suppliers, and the chip supplier's roadmap dictates how fast models scale. What connects all of this isn't a single trend line.

00:09:52 It's the same structural pressure showing up in different registers. Capital markets are pushing back on infrastructure spend. Governments are putting moratoriums on data center permits. Talent is deciding Silicon Valley isn't worth the visa risk. Sanctions are forcing domestic chip development.

00:10:12 And adoption economics mean that even as models improve, companies paying for tokens still aren't making the revenue catch up. The local read here is simpler than the headlines suggest. We're in the phase where building AI infrastructure costs more than anyone can agree will eventually pay for it — but everyone needs to be there anyway.