Alphabet raised its 2026 capital spending guidance to $195 billion-$205 billion, up from $180 billion-$190 billion, even as free cash flow swung negative for the first time, the Korean blogger Meru writes in a same-day breakdown of the earnings call. He had flagged this report as the week's most important event because it would be the first real test of whether hyperscalers keep funding the AI buildout, and the numbers answered clearly, they are spending more, not less. First-half capex now totals $80.6 billion, $35.7 billion in Q1 plus $44.9 billion in Q2, meaning the second half needs $57.2 billion to $62.2 billion per quarter to hit the new guidance, more than the first half's pace. On the earnings call, Google's CFO said the company plans to lease more third-party computing capacity in the third quarter as a bridge while it builds out its own data centers, and reiterated that Google remains compute constrained with very strong AI demand. Cloud order backlog grew from $462 billion in Q1 to $514 billion in Q2, an 11.3 percent rise Meru calls solid evidence that memory demand keeps building, though he notes bulls will call the record backlog reassuring while bears will call the 11 percent sequential growth a slowdown from the prior quarter's near-doubling. The catch is cash, operating cash flow of $39.1 billion minus $44.9 billion of capex left free cash flow at negative $5.8 billion, down $11.2 billion from a year earlier. Meru's one-line read: Google shareholders may be nervous about that negative number, but for Samsung Electronics and SK Hynix, a customer that keeps spending more is not bad news.
Why it matters · Alphabet is the bellwether for whether hyperscaler AI spending keeps accelerating, and its raised guidance despite negative free cash flow is the clearest signal yet that Samsung and SK Hynix's memory demand isn't about to slow.
Worth asking · Google just raised AI capex guidance while free cash flow went negative for the first time. Is spending through a cash flow hit a sign of confidence in AI demand, or the first crack in hyperscaler discipline?