Announced vs. Shipped: Nvidia's 8-Month GPU Gap, Tracked
"Announced" and "shipped" get used interchangeably in AI chip coverage, and that's a problem if you're trying to time anything off the news. So this post counts it directly: across Nvidia's last five GPU generations, how long does the gap actually run between a spec announcement and the first real shipment?
KEY TAKEAWAYS
1. Across H100 through Rubin, the announcement-to-shipment gap has averaged 7.8 months, ranging from 6 to 9 months.
2. Rubin first appeared as a name on a roadmap slide in June 2024. Its planned shipment is fall 2026 — about 28 months later.
3. Memory-maker earnings move ahead of GPU shipment headlines, because sample submission and customer qualification happen well before a GPU ships.
Counting five generations
Using only publicly verifiable dates, the pattern is consistent: the quarter promised at announcement is rarely the quarter delivery actually happens in.
H100 was guided for Q3 2022; partner systems arrived in October, and DGX systems slipped to Q1 2023. H200 was guided for Q2 2024; real deliveries didn't start until Q3. B200 is the most-cited case — announced March 2024, a mask revision was reported that August, and GB200 mass production landed in December.
Roadmap, spec announcement, and shipment are three different events
Rubin is a clean illustration. There are three distinct moments: the name "Rubin" first appeared on a roadmap slide at the June 2024 Computex keynote — at that point it was a plan, not a product. In January 2026, Nvidia announced it had entered mass production at CES, disclosing specs. The actual shipment is planned for fall 2026.
Headlines cover all three as "Nvidia unveils next-gen GPU." The same product becomes news three separate times over two years. Treat each mention as a fresh data point, and your read on timing drifts.
What this means for reading the memory side
From the memory makers' vantage point, the order is: submit samples, pass customer qualification, enter mass production — and only then does the GPU ship. By the time a GPU shipment headline runs, memory makers have already moved through several quarters of that process.
That's why "16-layer HBM ready for mass production," covered in part 2 of this series, tends to land before the GPU it feeds into ships. The reverse also holds: when a GPU's timeline slips, memory inventory builds up first. Part 1 of this series covered exactly that — TrendForce cut Rubin's share of 2026 HBM allocation from 29% to 22% while raising Blackwell's HBM3E share from 61% to 71%.
What I actually watch
| Checkpoint | Why it matters |
|---|---|
| Whether Rubin holds its fall 2026 date | On schedule means a 9-month gap; a slip sets a new record for this series. |
| Nvidia's quarterly inventory and prepayments | Inventory and supply-chain prepayments move ahead of shipment when a launch is imminent — a more honest signal than announcements. |
| Customer qualification news from the three memory makers | This indicator leads GPU shipment; a delay here raises the odds of a GPU delay too. |
| Server OEM shipment guidance | Server makers' commentary is often closer to the real timeline than Nvidia's. |
Value chain read-through
| Stage | Who moves | Indicator to track |
|---|---|---|
| Sample submission | Three memory makers | Sample supply announcements, spec revisions |
| Customer qualification | Three memory makers | Qualification pass reports, resubmissions |
| Memory mass production | Three memory makers | Mass-production announcements, layer count |
| GPU shipment | Nvidia / server OEMs | Inventory, prepayments, shipment guidance |
| Revenue recognition | Cloud providers | Capex plans |
Risks to this view
• Shipment dates here are approximations built from public reporting. Initial low-volume shipment and full-ramp shipment are defined differently product to product, so month-level precision shouldn't be assumed.
• The sample size is five generations. That's thin for projecting a 7.8-month average onto whatever comes next.
• The gap could shrink instead of widen. A more mature supply chain tends to close the distance between announcement and shipment.
• The opposite is also plausible: rack-scale integrated products require more validation, which could stretch the gap further.
Starting with the next post, this series shifts to a recurring quarterly feature: tracking HBM exposure across Samsung, SK hynix, and Micron — three DRAM makers whose HBM weighting looks nothing alike.
Sources: Nvidia press materials (Mar 2022 / Nov 2023 / Mar 2024 / Mar 2025 / Jan 2026 / May 2026); HPCwire (Mar 2022 / Sep 2022); Computex 2024 keynote; TrendForce (Sep 2024 / Apr 2026); related press coverage.
Disclaimer: This post is for informational and educational purposes only. It does not constitute investment advice or a recommendation to buy or sell any security. All investment decisions are your own responsibility.
Comments
Post a Comment