Fifty-one billion dollars of capex in a single quarter, and free cash flow comes back nine billion in the red. That's Alphabet. Not some neocloud. And Microsoft, the same afternoon, threw off cash. This is The Data Center Daily for Thursday, October 29th. Today: Microsoft and Alphabet report, and the cash-flow split between them. CAISO's board approves its large-load rules, with PJM due back at FERC today. Batch Zero heads toward litigation in Texas. And Latitude Media on cracked turbines in the onsite gas rush.
From Particle's roundup of Wednesday's earnings:
Microsoft and Alphabet reported on Wednesday that cloud revenue surged, with Azure at about $33.2 billion and Google Cloud at roughly $30.7 billion, reflecting rapidly rising enterprise spending on AI compute.
For this audience, the line that matters is capex. Particle puts Alphabet's spend at roughly 51.4 billion dollars for the quarter, leaving a free cash flow deficit of about 9 billion. Microsoft generated roughly 18.6 billion in free cash flow. Compare that with Vantage Markets' preview, which had Alphabet's second quarter at 44.9 billion of capex and negative 5.9 billion in free cash flow. So the spend went up and the hole got deeper. Google Cloud is growing, sure. Not fast enough to pay for the concrete. Going in, Alphabet had already raised its 2026 capex range to 195 to 205 billion, and finance chief Anat Ashkenazi called the company supply constrained. Microsoft had guided to more than 50 billion for the September quarter, and Amy Hood said customer demand continues to exceed available capacity. The market read it the obvious way. Per Particle, Alphabet shares fell sharply on weaker ad growth and cash burn, and Microsoft rose. Investors love the buildout right up until it prints as a negative number. Particle's summary doesn't carry updated capex guidance from either company, so that's the figure to pull from the filings. Amazon reports today after the close. In an a16z interview published October 23rd, Andy Jassy said AWS capacity for 2026 is sold out, 2027 largely sold out, and a good chunk of 2028 as well, with 2026 capex around 220 billion dollars. Sold out is a demand claim. What I want tonight is whether that 220 moves again, and what he says about power.
The California ISO, in its release after Wednesday's board vote:
A key element of the Board-approved proposal, referred to as a “flexible interim load interconnection,” would enable large loads to interconnect quickly to the transmission system with enforceable flexibility determined by grid conditions and constraints until associated network upgrades are complete.
We previewed this vote yesterday. The board approved it, and CAISO will file with FERC by its November 16th deadline. FERC then decides whether it satisfies the June show-cause order on large loads. Flexible interim interconnection. Translation: plug in before your upgrades are built, and accept getting cut when the grid says so. I'm for it, as long as enforceable means enforced. The release spells out the teeth: hourly forecasts and telemetry, plus communication protocols so CAISO and the transmission owner can remotely disconnect a large load to protect reliability. Co-located load and generation would file parallel application information, so CAISO sees the whole arrangement. And CAISO says it will keep working with NERC on reliability standards for computational loads. Remote disconnect, written down. After PJM told FERC that about 3,900 megawatts of large load jumped to backup generation on July 22nd, that's not a footnote. Generation serving large loads at separate but nearby sites goes to future policy work. The release doesn't mention PG&E's objection to the 50 megawatt definition. Also at FERC today: PJM said it would refile its Reliability Backstop Procurement, the one-time buy of roughly 6.8 gigawatts that FERC suspended on September 29th, largely over who pays. Energy News Beat reports PJM's preview would allocate costs by each zone's share of large-load growth. The zones that bring the data centers pay for the plants. Right instinct. Let's see whether it survives the actual filing.
The Power Grab Texas morning briefing, on Wednesday:
Staff told the commission it "should expect between 12 and 20 additional such complaints in the coming weeks," and warned that, unless they're managed efficiently, "the available resources of ERCOT, Staff and the Commission will be overwhelmed."
Yesterday we had CenterPoint's count of roughly 12 gigawatts of Houston load still sitting in Batch Zero. This is the fight over who stays in it. On Monday, October 26th, Skybox Datacenters became the fifth developer to file a complaint against ERCOT at the Texas PUC, over its Wichita Falls campus. The megawatts are redacted in the public copy; earlier local reporting put the campus at about 800. And look at how they lost their spot. ERCOT's September 3rd provisional classification put Skybox in studied load instead of base load, because its approved stability study left out 19 other large loads ERCOT found afterward. ERCOT says Oncor never submitted the updated study by July 10th. So the developer gets bumped over its utility's homework. Skybox, represented by Baker Botts, wants the commission to reverse ERCOT and restore base-load status. The four earlier complainants are Medina County Power, Rowan Stampede, Mesquite H2 and ACS DC Hicks Field Road, all seeking expedited schedules. On Tuesday, administrative law judges abated those four cases and paused Skybox's until November 10th. Staff wants one common process set at the November 5th open meeting. If another 12 to 20 are coming, that meeting decides how two dozen developers argue their place in line. For them, that could matter more than the December 10th audit report. One more data point from the same briefing: a University of Houston poll of 2,000 likely voters found 99 percent support making data center companies pay for the grid upgrades they need. Ninety-nine percent. You can't get ninety-nine percent of Texans to agree on brisket.
Latitude Media, in its AI-Energy Nexus newsletter:
We’ve been hearing stories about large power providers and EPCs fielding a growing volume of calls from data center developers that are blowing gas-fired turbines, and need help fixing mechanical failures because they lack their own operations and maintenance expertise.
The backdrop: Cleanview estimated this month that 110 gigawatts of onsite gas projects are underway, up from 20 last summer. The failures Latitude rounds up include cracked turbines at small U.K. projects and at xAI's Colossus campus in Memphis, per Bloomberg, where xAI added Tesla batteries to absorb power swings. Crusoe's Stargate site in Abilene ran into problems with backup turbines and cooling, per The Information. And Vantage's VA2 site in Loudoun County had what was called a critical power failure at its onsite gas plant this summer. Nearly 24 hours on backup diesel. The backup plant needed a backup, and the neighbors got the noise and the exhaust. The cause, per Latitude's reporting: AI workloads can swing by megawatts within milliseconds, and that stresses equipment. Developers are pairing gas with storage, sometimes with limited or no grid connection, plus UPS systems inside the hall. So now you're running a power plant, a battery plant and a data center, with a team that's done maybe one of those. TerraFlow Energy's CEO, Jon Parella, said it plainly: you need mechanics who can fix and maintain gas engines and service them every year. Then there's fuel. Latitude points to Anthropic's S-1, which we covered Tuesday, flagging fuel price shocks and pipeline constraints for behind-the-meter gas, and saying clean alternatives may not arrive in the quantities or on the timelines it needs. Speed to power is the pitch. Uptime is the invoice.
If you operate onsite generation and think Latitude got something right, or wrong, tell us. The address is datacenterdaily at lantern podcasts dot com. Turbine war stories especially welcome.
Tonight it's Amazon's capex number, and whether PJM gets its backstop filing in on time. We'll pick up both tomorrow, Friday. You've been listening to The Data Center Daily, a Lantern Podcast. Thanks for spending part of your morning with us.