Friday's edition shipped Nvidia's Q1 FY27 revenue at "$75 billion." The actual number is $81.6 billion. [1] The data-center segment was $60.4 billion. The Q2 guide is $91 billion. The board authorised a fresh $80 billion no-expiration share buyback on top of $38.5 billion still left from the prior authorisation, with $20 billion of the prior authorisation already executed in Q1. The quarterly dividend was raised from $0.24 to $0.25. The Vera CPU architecture was introduced with a stated $200 billion total addressable market. The paper's Friday brief understated the print's headline by $6.6 billion and the data-center line by $14.8 billion; the correction belongs in this body, not in a footnote.

The arithmetic mistake was the $75 billion figure read as a data-center number when it was a different segmental cut. The actual data-center compute line is $60.4 billion; the broader data-center category that Reuters and Bloomberg both reported includes networking, which alone was $14.8 billion — up 199 percent year-over-year and 35 percent sequentially. [2] Networking is now larger than the entire company's revenue line in any quarter before fiscal year 2024. The composition is the story Friday missed. The company is not just selling GPUs; it is selling the interconnect that lets the GPUs talk to each other inside a hyperscaler data-centre, and the interconnect line is growing twice as fast as the chips it connects.

The Q2 guide is the Saturday tape. $91 billion of revenue for the quarter ending in late July would put Nvidia on roughly a $364 billion annualised run-rate by mid-summer — a figure that exceeds the combined annual revenue of the next four largest US semiconductor companies. [3] The guide does not assume material new product launches; it assumes the existing Blackwell ramp continues at the pace observed in Q1, with Hopper revenue flat and Grace Blackwell systems shipping at increasing volume to the Stargate and SoftBank-Oracle-Anthropic build-outs that consumed most of the Q1 data-center allocation.

The Vera CPU disclosure was the second-day story Reuters and the analyst desks ran with. The Vera architecture — Nvidia's next-generation CPU paired with Rubin GPUs in late-2026 servers — was described in the earnings call as addressing a "$200 billion total addressable market" inside the existing data-center estate, primarily by displacing x86 incumbents Intel and AMD in AI-paired server roles. [3] The TAM figure is the company's, not an independent analyst's, and the announcement frame is forward; what matters for the Q1 print is that Nvidia is no longer selling Blackwell against a CPU it does not control. The Vera roadmap closes the only architectural seam Blackwell shipped with.

The buyback geometry is the part that surprised the second-day analyst tape. $80 billion authorised with no expiration, on top of $38.5 billion still left from the prior authorisation — gross authorisation capacity of $118.5 billion — is roughly equal to the company's total revenue across fiscal years 2022 and 2023 combined. The $20 billion already executed in Q1 is the kind of capital return that frontier-lab compute customers would, in a different macroeconomic frame, prefer to see absorbed by capacity build rather than share retirement. The market did not punish the announcement; Nvidia closed Friday at $215.42, down 1.86 percent on no apparent news catalyst, with sell-side firms broadly raising price targets in the wake of the call. [3]

The chip-supply chain reaction was the predictable cross-read. Samsung and SK Hynix — both Nvidia high-bandwidth-memory suppliers — saw share drops as investors took profits on the indirect-exposure trade. [4] The narrower point: Nvidia's earnings call now moves the entire memory complex's pricing in real time, which is itself an indication of how much of the global semiconductor demand curve is now downstream of a single company's guidance.

The Anthropic-shaped piece of the print is the one this paper has been tracking through the week. Nvidia's earnings call disclosed it holds private AI stakes in Anthropic, OpenAI, and xAI at cost — values updated only on observable events like new financings. [5] The Anthropic round closing at a $900 billion valuation this weekend is exactly the "observable event" that should produce a Nvidia mark-to-model adjustment on its Anthropic stake at the next quarterly close. The compute scale Anthropic's $45 billion SpaceX contract anchors is the same scale Nvidia's $60.4 billion Q1 data-center line monetises.

Friday's number was wrong, and the wrong number understated a company that has now booked $81.6 billion of revenue in a single quarter, guided $91 billion for the next, authorised $80 billion of fresh buyback, and committed publicly to a $200 billion CPU TAM. Each of those four numbers is larger than the entire AI sector's annual revenue from three years ago. The paper's job is to print the right ones.

The right ones are above.

-- DAVID CHEN, Beijing