GTC Berlin Recap: Vera Rubin Full Production and the Spillover into Gamer Hardware#

Zixing | October 22, 2026

Hello everyone, I’m Zixing.

GTC came to Europe for the first time this year — it was held in Berlin, Germany on October 20–22, and Jensen Huang took the keynote stage locally on October 21. Two days earlier, in the article where NVIDIA officially framed the GPU shortage, I had put “whether GTC Berlin would mention supply arrangements for the gaming business” on my watch list. Let me say the awkward part upfront: that article covered the supply-chain market and the official framing; this one only breaks down the technical announcements of the keynote itself, and I will not repeat a single market argument from it. The two things should be kept separate — one is where the market goes, the other is where the technology is headed.

This piece has exactly one landing point: what does a datacenter-focused keynote have to do with those of us sitting at a PC playing games. I split the keynote’s information into two threads: one covers the technical essentials of Vera Rubin’s full production and the new datacenter baseline; the other traces how these infrastructure advances spill over toward gamer devices — the memory and compute specs of the local AI device line, which I have been tracking hands-on for the past six months, finally got a new official anchor.

As always, every figure comes from post-event public reporting and NVIDIA’s official statements; which items are company self-reports and which have not been independently re-tested are listed item by item at the end.

I. Three Hard Metrics of Full Production: 5 Minutes, 100% Liquid Cooling, Cable-Free#

The heaviest sentence of the keynote: the Vera Rubin platform is in full production. For datacenter customers that is a delivery signal; for someone with my teardown habits, what really matters are the three engineering metrics behind full production — they set the deployment tempo of next-generation compute.

First metric: assembly time compressed from about 2 hours to under 5 minutes. Assembling a compute node on the previous-generation Grace Blackwell platform took about 2 hours; the Rubin platform compresses that to under 5 minutes, a more-than-24x speedup. The lever is cable-free design — nodes require no cables, paired with full liquid cooling. A 24x faster assembly is not showing off; it directly determines how quickly the statement “an AI factory only creates value once it starts producing tokens” can be cashed in: every day of slow installation is a day those hundreds of megawatts in the building sit idle.

Second metric: 100% liquid cooling. The previous platform’s liquid-cooling coverage was 80%, with the remaining fifth of nodes falling back to air cooling; Rubin goes all the way to 100%. The direction of change is clear — thermal design in server rooms is converging on full liquid cooling, because air cooling no longer suffices at high power density.

Third metric: 72 GPUs interconnected into one giant GPU. Rack-scale interconnects like NVL72 bind 72 GPUs into a single logical unit; high-speed in-rack interconnect replaces cross-machine communication overhead. This extends the “the rack is the computer” idea, and on the deployment side it means single-rack power density and cooling requirements rise in step — which is precisely the premise that makes 100% liquid cooling hold up.

On the performance side there are two more company-stated figures; here they are, re-testing discussion comes at the end:

When reading the 30x-per-megawatt number, hold on to the words “per megawatt” — it is not a single-card performance comparison but throughput obtained per unit of electricity. That framing only makes sense inside the economics of an AI factory; next section expands on it.

A table makes the generational change more intuitive:

This table’s value is not in any single row but in how the three rows interlock: cable-free design is the precondition for the 24x assembly speedup, 100% liquid cooling is the precondition for high-power-density racks, and rack-scale interconnect is the precondition for “30x per megawatt” — the three metrics form one co-designed package, and reading any of them in isolation understates the engineering significance of this production milestone. Jensen Huang summed up the approach as “extreme co-design” and offered the 11-year cumulative 1,000,000x AI performance gain as the long-term figure.

II. DSX POD: Shipping AI Factories Like Containers#

If the Rubin platform is “built fast,” DSX POD is “deployed fast” — in my view the announcement with the biggest impact on industry tempo at this event.

The traditional AI datacenter build sequence is: site construction → rack placement → cabling → integration → testing, several stages run in series, easily taking months. DSX POD’s idea is to move that whole sequence into the factory: 7 Rubin liquid-cooled racks, together with about 17 kilometers of Spectrum-X cabling, are pre-assembled into a standard shipping container at the factory, running on 800 VDC power, with integration, cabling, and testing completed before delivery. On site, the container lands, connects to power and water, and goes live.

Comparing the two deployment-cycle figures:

I use the official target of “as fast as 14 days,” but one premise must be stated: on-site utilities (power, cooling) are already in place. In other words, the 14 days compress rack-level integration and testing time; site construction and grid connection still take as long as they take — don’t read “as fast as 14 days” as “a datacenter built in 14 days.” DSX POD deliveries by global partners are planned to begin in early 2027.

Per the model announced at the keynote, the DSX POD delivery chain breaks into four steps:

1. Factory pre-assembly: 7 Rubin liquid-cooled racks plus about 17 kilometers of Spectrum-X cabling are fitted into a standard container, with the 800 VDC power architecture laid out at the same time;
2. Factory integration and testing: integration, cabling, and testing are all completed before delivery — moving the “on-site manual work” that is most error-prone in traditional server rooms up onto the production line;
3. Container shipping and landing: the container arrives on site, and the server room only needs to provision power and cooling interfaces;
4. Go live once utilities are connected: when on-site utilities are ready, take the “as fast as 14 days” path; if not, it reverts to the traditional cycle — that is exactly what the premise means.

Three supporting details are worth noting:

1. Closed-loop liquid cooling supports coolant up to 45°C — the higher the inlet-temperature ceiling, the more the system can rely on free cooling and the shorter the compressor’s runtime window, improving overall energy efficiency;
2. In climate-suitable regions, outdoor dry coolers can reject heat to reduce water consumption — a signal that liquid cooling is moving from “water guzzler” toward “water-saving form”;
3. Vertiv’s 2.3MW CoolChip coolant distribution unit, among others, earned DSX Ready status — the supplier ecosystem is building a certification system around this pre-fabrication standard, and the supply chain division of labor is getting finer.

The 800 VDC power item is, in my view, the design in the whole DSX scheme that best signals direction: once rack power density climbs, the conversion stages and copper losses of AC distribution start eating efficiency, and high-voltage DC cuts out part of the conversion chain. This is datacenter-level power architecture evolution.

III. AI Factory Economics: $50 Billion, $40 Billion, and a $10 Billion Loss from a Three-Month Delay#

Jensen Huang gave a set of AI factory economics in the keynote; I reproduce them as-is (all company figures):

Read together, this table explains why the engineering metrics in Section I deserve scrutiny: with $50 billion of capex sitting in a building, every month of delay is a revenue hole on the order of a billion dollars, so the 24x assembly speedup, the 14-day deployment, and the 30x-per-megawatt throughput are essentially three variables of the same financial model — get tokens flowing earlier, and get more tokens per kilowatt-hour. His line “we are essentially industrializing the mass manufacturing of supercomputers” says it outright.

On the European side: NVIDIA has committed nearly $10 billion in investment in Europe with more than 12,000 partners; European carriers such as Orange Business, Fastweb, Swisscom, Telefónica, and Telenor are expanding sovereign AI factories. Google Cloud announced that its Rubin-based A5X instances are planned for full availability in November, with related infrastructure supporting cross-site scaling to up to 960,000 Rubin GPUs — read that as the upper bound of a cloud provider’s stated purchase intent for the Rubin platform, a planning figure rather than deployed capacity.

IV. Supply-Chain Expansion List: Six Names and Five Numbers#

“Full production” is not a slogan; behind it is simultaneous expansion across the supply chain. The six partners and their numbers announced at the keynote (all as stated at the event):

Two more aggregate figures: total test capacity doubles from 2026 to 2027; and Jensen Huang’s characterization of demand: “demand is extraordinary, capacity is the only constraint.”

A separation statement must be inserted here: the expansion figures above are datacenter supply-chain actions announced at the keynote and belong to a different track than consumer GPU market analysis — for the market framing, root causes, and price data, see the October 20 article; this piece does not repeat them. The only inferences I can draw from the fact cards are directional, and both are labeled as inference:

• Inference 1 (direction, not fact): large-scale expansion of packaging, substrate, and test capacity means advanced packaging and test lines are being expanded with datacenter products in mind. How newly built capacity gets allocated will affect how much advanced capacity consumer products can claim in the future — but the keynote gave no ratios or timelines, and any specific judgment would be fabrication, which I won’t do.
• Inference 2 (direction, not fact): if “capacity is the only constraint” keeps holding, the pace at which expansion lands is the first signal for watching supply changes — harder evidence than any channel rumors.

Console players care about AI features trickling down; this piece’s fact cards do not cover console platforms, so I won’t expand on it; the PC-side trickle-down evidence is in the next section, which is actually the part of this keynote players should read most closely.

V. The Local AI Device Line: The Real New Anchor for Gamer Desktops#

The most directly informative part of this event for players is not the datacenter but the three nodes of the local AI hardware product line. Official specs, organized into a table:

Each node carries one line of meaning for players:

RTX Spark laptop — “personal agents” finally have a concrete hardware form. The 128GB unified memory is the key item: the barrier to running large-parameter models locally has never been compute but memory capacity, and a unified-memory architecture lets CPU and GPU share one large memory pool. The 1 petaflop AI compute rating means it has headroom for local inference. Model, OEM partners, and release date are all unannounced, so this one goes on the watch list for now.

The DGX Spark 64GB goes on sale tomorrow (October 23), and an 8-unit cluster reaches 1TB of unified memory. For serious local-deployment players, this is the first consumer-reachable device that is “a cluster out of the box” — a single 64GB unit runs small-to-mid models, and once 8 units are interconnected, a 1TB memory pool can hold model weights far beyond a single machine’s capacity.

DGX Station is the ceiling of the local device line: 748GB coherent memory, up to 20 petaflops, Windows version in December. The Windows version matters for the ecosystem — the compatibility and deployment cost of local AI toolchains on Windows directly determine whether ordinary players can actually use it.

Lining up the memory capacities of the three devices reveals a clear ladder: RTX Spark laptop 128GB → DGX Spark single unit 64GB (1TB across an 8-unit cluster) → DGX Station 748GB. The meaning of this ladder: the upper bound of usable local large-model size now has stepwise tiers from laptop to workstation. Anyone running local models knows capacity tiers become the bottleneck before peak compute does — if the weights don’t fit in memory, all the spare compute just means quantizing down a notch. With the 128GB / 748GB / 1TB-cluster capacity anchors officially on the table, the old question “how big a model can I run locally” now has three matchable answers.

Two supporting software and security items:

• Open-source local coding model Recursive Flint: a 40B MoE with 2.5B active parameters, reaching over 220 token/s on RTX Spark, weights open-sourced under Apache 2.0. This is the first time the local device line has an official built-in “run it and see” demo workload — the 220 token/s figure is the official number; I will re-test it once I get hardware.
• OpenShell reference architecture: each agent runs in its own kernel-level sandbox, least privilege enforced at syscall granularity, and a detected privilege violation cuts its connection to the AI model within milliseconds. If the first three items answer “can an agent run locally,” OpenShell answers “would you dare let it run” — for players planning to put agent plugins on their gaming or daily machines, the sandbox isolation mechanism is a security baseline worth learning in advance. Breaking down its three layers: the kernel-level sandbox contains the blast radius, so one compromised agent does not touch the host system; syscall-granularity least privilege shrinks the attack surface, every system capability an agent gets requires individual authorization; millisecond cutoff limits damage, disconnecting the agent from the model first and talking rollback later. Three layers, three jobs — this rubric can be applied directly to any future “AI security tool” review.

A batch of open-source model families also shipped (Nemotron 3.5 Super VL/Bolt language models, Nemotron Retriever for retrieval, Kumo for structured data prediction, Cosmos 3.5 world models, Alpamayo 2 Edge for in-vehicle inference, BioNeMo RFD4-Proteina protein design in collaboration with the University of Washington Baker Lab, and the Ising series for quantum); they are only loosely related to gamer desktops, so I list names only without expanding.

VI. Other Announcements Loosely Related to Gamer Hardware: For the Record#

A complete post-event recap should also archive the other two blocks of the keynote that are only loosely related to gamer hardware — they illustrate the same trend: AI is penetrating from cloud into all kinds of edge devices.

Physical AI and robotics: NVIDIA and Agile Robots unveiled the Agile ONE SN humanoid reference design with 31 degrees of freedom, carrying Isaac GR00T and Jetson Thor, with the body planned for early 2027. Its single point of contact with gamer hardware — Jetson Thor is an edge compute device, and the production tempo of such reference designs is one corroboration for watching the “price curve of AI compute trickling to the edge.” That is all; it is not a forecast for consumer products.

Automotive and telecom: Mercedes-Benz, Volkswagen Commercial Vehicles, Einride, Lucid, and others showed L3/L4 and Robotaxi use cases on the Hyperion platform, and Volvo Cars adopts the Hyperion and Alpamayo software stacks; on telecom, NVIDIA is working with Nokia and multiple carriers to turn telecom networks into distributed intelligence platforms, and the first commercial call has been completed on the T-Mobile network. Both are industry-level partnerships with limited direct meaning for players; archived for future tracking.

VII. A Player’s Watch List: What Matters and What Doesn’t#

Having taken the whole piece apart, let me draw the “relevant / not relevant” boundary clearly so clickbait doesn’t drag you around. First a quick-reference table of key dates, all as announced at the keynote:

Directly relevant to players:

1. The three dates of the local AI device line (DGX Spark on sale October 23, DGX Station Windows version in December, RTX Spark laptop TBD) — concrete milestones you can put on a purchase-waiting list;
2. The figures 128GB unified memory, a 1TB cluster memory pool, and 748GB coherent memory — they set the capacity baseline for local inference devices over the next two to three years, and the feasibility ceiling of game-side local AI features (NPC dialogue, voice, image-generation apps) follows these numbers;
3. Test capacity doubling from 2026 to 2027 and the partners’ production dates (April 2027, mid-2027, end of 2028) — public anchors for watching the supply turning point.

Not directly related to players — don’t over-read:

1. 800 VDC power and 45°C closed-loop liquid cooling are facility-grade architectures — they change the build and operating cost model of datacenters and have no direct mapping to gamer PSUs or all-in-one coolers. Seeing the word “liquid cooling” and dragging datacenter thermal design onto your desktop is the most common misreading of this news;
2. DSX POD’s 14-day deployment, the 960,000-GPU Rubin plan, and the $50B/$40B economics — all enterprise figures; they shape the macro rhythm of compute supply and are not a direct prediction of any consumer product;
3. The keynote contained no consumer GeForce gaming card information at all — no specs or timeline for a next-gen gaming architecture, and Berlin coverage included nothing on DLSS evolution or a GeForce roadmap. Any headline that leaps from “Rubin is in production” to “here’s what next-gen gaming cards will do” is adding drama beyond the facts.

One line to sum up the boundary: this keynote rewrote how AI factories are built and run; gamer desktops are affected indirectly, through two narrow openings — the local device line and long-term capacity allocation. The first-hand data through the narrow opening is the DGX Spark going on sale October 23.

VIII. Figures and Unconfirmed Items#

As always, let me lay out the quality of the evidence before closing:

• Company figures without re-tests: 30x throughput per megawatt, 1,000,000x over 11 years, deployment as fast as 14 days — all from NVIDIA’s keynote, with no independent re-test data; this article presents them as relayed claims;
• Two deployment-cycle figures coexist: The Paper gives “about 6 months → as fast as 14 days,” Xin Dongxi gives “4–5 months → about 1 month”; this article adopts the official target and notes the “on-site utilities ready” premise;
• Not announced: no official information this time on a consumer Rubin-architecture GPU (the rumored next-gen RTX) timeline; the RTX Spark laptop’s model/OEM/release date unannounced; no DLSS evolution or GeForce roadmap content appeared in Berlin keynote coverage;
• Inference labels: the two inferences in Section IV are directional analysis, not fact; all expansion figures are as announced at the keynote;
• Sources: this article’s factual basis comes from post-event public reporting by The Paper (republished on Toutiao), Xin Dongxi (NetEase account), a Soochow Securities commentary (Tencent News), Toolnavs AI Briefing, and NVIDIA’s official site; cross-confirmed points are noted, and single-source figures are labeled.

Final Thoughts#

Having written this, my read of the event fits in one sentence: NVIDIA has turned “building computers” into an assembly line, and for the first time made the line’s takt parameters (5-minute assembly, 14-day deployment, 24x speedup) the star of the keynote. Datacenter readers look at capacity; I’m watching something else — when $50-billion-class factories roll out at one container per 14 days, the line expansions in advanced packaging, substrates, and testing will, at some link, branch off a tributary toward the consumer market. When that tributary appears and in what form, this event didn’t say; I’ve only left the two directions tagged “inference” in Section IV.

The more certain thing is right in front of us: tomorrow the DGX Spark 64GB goes on sale, the DGX Station Windows version follows in December, and the local AI device line has its first complete price-and-spec ladder from laptop to workstation, from 128GB to a 1TB cluster. For players who want to run local models on the desktop, this year-end buying window is clearer than ever — as for when it’s my turn to get one for hands-on testing, that waits for the day the RTX Spark laptop’s model is announced.

I’m Zixing — see you in the next hands-on piece.