Nvidia's Newest AI Hardware in 2026: Blackwell Ultra, Vera Rubin, RTX Spark and the Road to Feynman Explained

By late October 2026, Nvidia sells a lot more than graphics cards. It sells whole racks of computers, the cables and switches that join thousands of those racks together, the processors that sit next to the GPUs, the software that schedules the work, and, as of this autumn, the main processor inside a new class of Windows laptops. It also makes a robot "brain" module, a desk-sized AI workstation, and a palm-sized one.

This guide explains where all of that stands as of October 31, 2026. It covers the Blackwell and Blackwell Ultra generation that most of the world's AI runs on today, the Vera Rubin generation now going into cloud data centers, the roadmap Nvidia has published for 2027 and 2028, the networking and memory behind it all, the desktop and robotics products, the consumer RTX line, the software stack, power and cooling, prices where they have been reported, the China export situation, and what competitors are offering.

Two ground rules apply throughout. First, every performance figure here is Nvidia's own claim unless we say otherwise. These are peak numbers from spec sheets and keynotes, not independent benchmarks. Second, we mark clearly which products are shipping, which are announced but not yet shipping, and which are roadmap only. Where a spec could not be confirmed from a primary or reputable source, we left it out.

The short version

If you only read one section, read this one.

• Blackwell (B200, GB200 NVL72) and Blackwell Ultra (B300, GB300 NVL72) are the workhorses in AI data centers today. A GB300 NVL72 rack holds 72 Blackwell Ultra GPUs and 36 Grace CPUs, and Nvidia rates it at 1,080 petaflops of dense FP4 compute. Shipping.
• Vera Rubin is the new generation. It pairs the Rubin GPU (288 GB of HBM4, 22 TB/s of memory bandwidth, and 50 petaflops of NVFP4 inference compute per GPU, by Nvidia's figures) with the 88-core Vera CPU. The flagship Vera Rubin NVL72 rack is rated at 3.6 exaflops of NVFP4 inference. Nvidia said it was in full production in January 2026. Microsoft Azure was the first hyperscale cloud to power up Vera Rubin NVL72 systems, and Oracle said on October 27, 2026 that the racks were generally available on its cloud. Shipping to cloud and large customers, ramping now.
• Groq 3 LPX is a rack of 256 specialized inference chips, built on technology Nvidia licensed from the startup Groq in December 2025. It is meant to sit next to Vera Rubin racks and speed up the token-by-token "decode" phase of chatbot answers. Announced for the second half of 2026.
• Rubin CPX, a long-context inference chip announced in September 2025, dropped off Nvidia's roadmap at GTC 2026. Analyst reports in September 2026 claim a redesigned version is back for 2027, but Nvidia has not confirmed that. Status uncertain.
• Rubin Ultra (2027, in the new 600-kilowatt-class "Kyber" rack) and Feynman (2028, with a new CPU called Rosa) are roadmap only.
• On the desktop, DGX Spark (a small box with 128 GB of unified memory and up to 1 petaflop of FP4) has shipped since October 2025. DGX Station with the GB300 desktop superchip and 748 GB of coherent memory is rolling out through partners.
• For consumers, the GeForce RTX 50 series launched in early 2025. The big 2026 news is RTX Spark, a single chip that combines a 20-core Arm CPU with a Blackwell GPU (6,144 CUDA cores) and up to 128 GB of memory. Nvidia says RTX Spark Windows laptops went on sale this autumn, starting at $2,599.
• In robotics, Jetson Thor delivers up to 2,070 FP4 teraflops (sparse) in a 40–130 W module, running Nvidia's open Isaac GR00T robot models and Cosmos world models.
• China: Nvidia's latest quarterly outlook assumes zero data center compute revenue from China.

A plain-English primer: what these machines actually do

Before the specs, it helps to know what the hardware is for.

A modern AI model, such as a chatbot, is a huge table of numbers called parameters or weights. Training the model means adjusting those numbers billions of times until the model's predictions get good. Running the model, which the industry calls inference, means pushing a user's question through those numbers to produce an answer one token at a time. A token is roughly a word or a piece of a word.

Both jobs come down to the same basic arithmetic: multiplying and adding enormous grids of numbers. A GPU (graphics processing unit) is good at this because it has thousands of small calculators working in parallel. Nvidia's GPUs also contain Tensor Cores, circuits built specifically for the matrix math at the heart of AI.

Three things limit how fast an AI computer runs:

1. Compute: how many calculations per second the chip can do. This is measured in FLOPS (floating-point operations per second). A petaflop is a thousand trillion of them; an exaflop is a thousand petaflops.
2. Memory: how much data sits right next to the chip, and how fast it can be read. AI chips use HBM (high-bandwidth memory), stacks of memory chips mounted beside the GPU. Memory capacity decides how big a model fits; memory bandwidth (in terabytes per second) often decides how fast tokens come out.
3. Interconnect: how fast chips can talk to each other. A frontier model is too big for one GPU, so it is spread across dozens or thousands. If the links between them are slow, the GPUs sit idle waiting for data.

One more concept matters for reading spec sheets: number precision. Numbers can be stored with many digits (FP32, 32-bit floating point) or very few (FP8, FP4). Fewer bits means less accuracy per number, but much faster math and less memory. Nvidia introduced a 4-bit format called NVFP4 with Blackwell. According to Nvidia's technical blog, it uses two levels of scaling factors so that accuracy stays close to FP8, often within about 1%, while cutting memory use by roughly 1.8x compared with FP8. Most of the eye-catching numbers below are FP4 or NVFP4 figures. They are not comparable to older FP16 or FP32 numbers.

Spec sheets also distinguish dense and sparse figures. "Sparse" numbers assume the model has been pruned so that some values are zero and can be skipped. Sparse figures are typically double the dense ones. Nvidia often quotes the sparse number. We note which is which where it matters.

Finally, the industry now talks about AI factories: data centers whose output is tokens. That framing explains why Nvidia increasingly quotes "tokens per megawatt" and "cost per token" rather than raw FLOPS.

Blackwell and Blackwell Ultra: today's workhorses

What Blackwell is

Blackwell is the GPU architecture Nvidia announced in March 2024 to succeed Hopper (the H100 and H200). Each Blackwell GPU is really two large silicon dies joined into one package. Together they hold 208 billion transistors, made on TSMC's 4NP process, according to Nvidia's developer blog.

The B200 is the standard Blackwell data center GPU. The GB200 "superchip" pairs two B200-class GPUs with one Nvidia Grace CPU (72 Arm Neoverse V2 cores). The GB200 NVL72 rack combines 36 of those superchips, which comes to 72 GPUs and 36 CPUs, all joined by Nvidia's NVLink so they behave like one giant GPU.

What Blackwell Ultra changes

Blackwell Ultra (B300, GB300) is a mid-generation upgrade that started reaching customers in 2025. It keeps the same transistor count, process, and two-die design. The changes are targeted at inference:

• 1.5x the dense NVFP4 compute of Blackwell (15 petaflops versus 10). Sparse NVFP4 is unchanged at 20 petaflops.
• 50% more memory: 288 GB of HBM3E versus 192 GB. Bandwidth stays at 8 TB/s.
• Faster attention: Nvidia doubled throughput of the special function units that handle a key step in "attention," the mechanism that lets language models weigh earlier words. Nvidia says attention-layer compute is up to 2x faster.
• More power: maximum board power rises from up to 1,200 W to up to 1,400 W.

Table 1: Blackwell vs Blackwell Ultra, per GPU (Nvidia figures)

Source: Nvidia developer blog, "Inside NVIDIA Blackwell Ultra." Some shipping B200 configurations are rated differently by partners; these are Nvidia's architecture figures.

Table 2: The rack-scale systems (Nvidia figures)

Source: Nvidia GB200 NVL72 and GB300 NVL72 product pages.

For eight-GPU servers, which many companies still prefer because they fit ordinary data center racks, Nvidia sells HGX B200 and HGX B300 boards. Both have 1.8 TB/s of NVLink per GPU. The B300 board carries 2.1 TB of total GPU memory versus 1.4 TB, and doubles networking bandwidth to 1.6 TB/s.

What Nvidia claims Blackwell does

Nvidia's headline claims for GB200 NVL72 compared with H100 are: up to 30x faster real-time inference for trillion-parameter language models, up to 4x faster training, and up to 25x more performance at the same power versus air-cooled H100 systems. For GB300 NVL72 it claims up to 50x higher "AI factory output" than Hopper on a DeepSeek-R1 reasoning test, with a footnote that the projection is subject to change. These comparisons use specific test setups chosen by Nvidia. Treat them as marketing ceilings, not typical gains.

On the independent side, Nvidia's August 2026 earnings release said Blackwell led every category in MLPerf Training 6.0. MLPerf is an industry benchmark suite run under shared rules, which makes it more comparable than vendor slides, though not every competitor submits to every test.

Vera Rubin: the 2026 generation

The naming

Nvidia names its architectures after scientists. This one honors astronomer Vera Rubin. The GPU is called Rubin and the companion CPU is called Vera, so the combined platform is Vera Rubin.

There was a naming wobble worth knowing about. At GTC 2025, Nvidia said it would start counting each GPU die as a "GPU." Under that scheme, a rack with 72 two-die packages would be called "NVL144." By CES 2026, Nvidia had reverted to counting packages, so the flagship rack is the Vera Rubin NVL72. Older articles that say "Vera Rubin NVL144" are usually describing the same 72-package rack.

The chips

At CES in January 2026, Nvidia described Vera Rubin as six chips designed together:

1. Rubin GPU
2. Vera CPU
3. NVLink 6 Switch
4. ConnectX-9 SuperNIC (network card)
5. BlueField-4 DPU (a "data processing unit" that offloads networking, storage, and security)
6. Spectrum-6 Ethernet switch

At GTC in March 2026, the Groq 3 LPU was added as the seventh chip (more on that below). Nvidia's GTC wrap-up described Vera Rubin as "seven chips, five rack-scale systems and one supercomputer."

Rubin GPU

Here is what Nvidia and its partners have published about each Rubin GPU:

• Compute: 50 petaflops NVFP4 for inference and 35 petaflops NVFP4 for training. The Register reported further figures: 17.5 petaflops FP8/FP6, 4 petaflops FP16/BF16, and 33 teraflops of FP64 vector compute. That FP64 figure is lower than Blackwell's 45 teraflops, a sign of how firmly Nvidia is tilting toward AI over traditional scientific computing.
• Memory: 288 GB of HBM4 at 22 TB/s. That is about 2.8x the bandwidth of Blackwell Ultra.
• NVLink: 3.6 TB/s per GPU, double the previous generation.
• Engine: a third-generation Transformer Engine with what Nvidia calls hardware-accelerated adaptive compression.
• Process: TSMC 3-nanometer class.

The "inference" versus "training" NVFP4 split is new. Nvidia's 50-petaflop inference number relies on features aimed at inference, so it is not directly comparable to Blackwell's dense 10 or 15 petaflops. Nvidia's HGX page labels it a sparse figure.

Vera CPU

Vera replaces Grace. Per Nvidia and The Register:

• 88 custom Arm cores (Nvidia calls them Olympus) with simultaneous multithreading and full Armv9.2 compatibility.
• Up to 1.5 TB of LPDDR5X memory per CPU, three times the GB200's.
• 1.8 TB/s NVLink-C2C link to the GPUs.

Nvidia also sells Vera as a standalone CPU. At Computex in June 2026, Jensen Huang said Vera was in mass production, with initial customers including Anthropic, OpenAI, xAI, Dell, Oracle, and CoreWeave. Oracle's October 2026 announcement says Vera CPUs are aimed at the CPU-heavy parts of "agentic" AI: running tools, executing code, and fetching data between model calls.

The Vera Rubin NVL72 rack

Table 3: Vera Rubin NVL72 at a glance (Nvidia figures)

Sources: Nvidia Rubin press release (January 2026), The Register's CES coverage, Nvidia's OCP blog.

Nvidia also highlights practical improvements: a cable-free tray design that it says makes assembly and servicing up to 18x faster than Blackwell; switch trays that can be swapped without powering the rack down; GPU health checks that run without pulling GPUs out of service; and confidential computing across the CPU, GPU, and NVLink domain, which Nvidia says is a first for a rack-scale system.

For customers who want ordinary eight-GPU servers, HGX Rubin NVL8 puts eight Rubin GPUs on a board with 2 TB of HBM4, 400 petaflops of sparse NVFP4 inference, and 24 TB/s of NVLink switch bandwidth. It pairs with x86 CPUs. Oracle says its planned HGX Rubin NVL8 servers use two 64-core Intel Xeon 6 CPUs and 4 TB of DDR5. A version with a single Vera CPU also exists. Even the eight-GPU board needs liquid cooling; Nvidia lists a 45°C inlet temperature.

Nvidia's claims for Vera Rubin

Compared with Blackwell, Nvidia claims:

• up to 10x lower inference cost per token for mixture-of-experts (MoE) models,
• 4x fewer GPUs to train MoE models,
• up to 5x greater inference performance (from CES coverage).

An MoE model splits itself into many "expert" sub-networks and activates only a few for each token. Most frontier models now use this design, which is why Nvidia frames its claims around MoE.

Where Vera Rubin stands now

The timeline, using Nvidia's own statements:

• January 5, 2026 (CES): Nvidia said Rubin was "in full production," with partner products arriving in the second half of 2026.
• March 16, 2026 (GTC): Nvidia said Microsoft Azure was the first hyperscale cloud to power up Vera Rubin NVL72 systems.
• August 26, 2026 (Q2 fiscal 2027 earnings): Nvidia said Vera Rubin was ramping into full production, with racks running at partners including CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure, and Nebius.
• October 27, 2026: Oracle and Nvidia said Vera Rubin NVL72 and Vera CPUs were generally available on Oracle Cloud Infrastructure, with plans to grow to multi-gigawatt scale.

So Vera Rubin is real and in customers' hands, but the volume ramp is still under way. Smaller buyers will mostly reach it through cloud providers first.

Groq 3 LPX and the Rubin CPX twist

Why inference got split in two

When a chatbot answers you, it does two different kinds of work. First comes prefill: reading your whole prompt at once. That step is compute-heavy and parallel. Then comes decode: generating the reply one token at a time. Decode is limited mostly by how fast the chip can read the model's weights from memory, because each new token needs another pass.

Nvidia's software, especially Dynamo, can run these phases on different hardware. That opened the door to specialized chips for each phase.

Rubin CPX: announced, then pulled

In September 2025, Nvidia announced Rubin CPX, a GPU for the prefill phase of very long prompts, such as a million tokens of code or an hour of video. Its press release gave these figures: 30 petaflops of NVFP4, 128 GB of cheaper GDDR7 memory instead of HBM, and 3x faster attention than GB300 NVL72. A combined Vera Rubin NVL144 CPX rack was to deliver 8 exaflops, 100 TB of fast memory, and 1.7 PB/s of bandwidth, which Nvidia put at 7.5x a GB300 NVL72. Nvidia also claimed $5 billion of token revenue for every $100 million invested, and said CPX would arrive at the end of 2026.

At GTC 2026, CPX was missing from the keynote. Wccftech, citing ComputerBase's reporting on comments by Nvidia vice president Ian Buck, said the product had been "delayed" and might reappear in the Feynman era. Then, on August 31, 2026, TF Securities analyst Ming-Chi Kuo reported that Nvidia had restarted CPX with HBM4 memory (168 GB per GPU in his account) and production expected in early 2027. Nvidia has not confirmed this. Treat Rubin CPX as unconfirmed for any date.

Groq 3 LPX: the decode booster

What Nvidia showed instead was Groq 3 LPX. In December 2025, Groq announced a non-exclusive licensing deal with Nvidia. Groq founder Jonathan Ross, who earlier co-invented Google's TPU, joined Nvidia along with Groq's president and other staff. CNBC put the value at about $20 billion. Neither company disclosed terms, and Groq continues as an independent company.

Groq's chips, called LPUs (language processing units), put memory directly on the chip as SRAM. SRAM is very fast but holds far less than HBM. That trade-off suits decode, where speed of access matters more than capacity. According to StorageReview's analysis of Nvidia's GTC materials:

• each LP30 chip has 500 MB of on-chip SRAM with 150 TB/s of bandwidth, about seven times the 22 TB/s of HBM4 on a Rubin GPU;
• each chip delivers 1.2 petaflops of FP8;
• a rack holds 256 chips in 32 one-unit trays, totaling about 128 GB of SRAM, with about 640 TB/s of scale-up bandwidth;
• the rack is fully liquid-cooled, built on Nvidia's MGX design, and hosted by x86 processors.

Nvidia pitches LPX as a partner to a Vera Rubin rack, not a replacement. The GPUs handle prefill and attention, while the LPUs handle the feed-forward and expert layers during decode. LPX can also generate draft tokens for "speculative decoding," which the GPUs then check. Nvidia claims up to 35x higher inference throughput per megawatt for latency-sensitive work than Vera Rubin alone. LPX is slated for the second half of 2026, initially for model builders and service providers. Announced; independent performance data is not yet available.

The roadmap beyond: Rubin Ultra (2027) and Feynman (2028)

Nvidia now publishes its roadmap years ahead and has committed to a yearly rhythm. Everything in this section is roadmap only. Specifications can and do change.

Rubin Ultra and the Kyber rack (2027)

At GTC 2025, Nvidia described Rubin Ultra as a package of four reticle-sized GPU dies with 1 TB of HBM4E, rated at 100 petaflops of FP4. ("Reticle-sized" means each die is as large as chipmaking equipment can print in one exposure.) It would ship in a new rack called Kyber, with 144 packages (576 dies) in one NVLink domain. That works out to about 15 exaflops of FP4 inference or 10 exaflops of training, per The Register's account of Nvidia's figures. The Register and other outlets reported Kyber at about 600 kW per rack.

Kyber stands its compute blades vertically, up to 18 per chassis, and connects NVLink switch blades through a cable-free midplane at the back. Coverage of the GTC 2026 roadmap associates Rubin Ultra and Kyber with NVLink 7 and a Groq LP35 chip that adds NVFP4 support. Kyber is to come in copper and co-packaged optical versions. Reporting has been inconsistent about whether to call the 2027 system "NVL144" (packages) or "NVL576" (dies). It is the same counting problem as before.

Feynman (2028)

Named after physicist Richard Feynman, this is the architecture after Rubin. Nvidia's GTC 2026 recap lists its pieces:

• a new Feynman GPU, which trade reports describe as using die stacking and custom HBM;
• a new CPU called Rosa;
• the LP40 LPU, which reportedly supports NVLink;
• BlueField-5 DPUs and ConnectX-10 (CX10) network cards;
• Kyber scale-up in both copper and co-packaged optics, plus optical scale-out.

Nvidia's NVLink Fusion page mentions a future roadmap with NVLink domains of up to 1,152 accelerators. No Feynman performance specifications had been published as of this writing.

At GTC 2026, Nvidia also showed a concept called Space-1 Vera Rubin, a system meant for data centers in orbit. Huang acknowledged radiation as an unsolved engineering problem. There are no specs; consider it a research direction.

Roadmap timeline

NVIDIA DATA CENTER ROADMAP (as stated by Nvidia, Oct 2026)

2024 |== Blackwell announced (GTC, March) -- B200, GB200 NVL72
     |
2025 |== Blackwell Ultra -- B300, GB300 NVL72 (shipping)
     |   DGX Spark ships (Oct) | Rubin CPX announced (Sep)
     |   NVLink Fusion, Spectrum-X / Quantum-X Photonics announced
     |
2026 |== VERA RUBIN -- Rubin GPU, Vera CPU, NVLink 6, CX-9,
     |   BlueField-4, Spectrum-6 (CPO); Vera Rubin NVL72, HGX Rubin NVL8
     |   [in production; on Azure, OCI, CoreWeave, Google Cloud, Nebius]
     |   Groq 3 LPX rack (2H 2026, announced)
     |   Rubin CPX dropped from roadmap (status unconfirmed)
     |
2027 |-- RUBIN ULTRA (roadmap) -- Kyber rack, ~600 kW,
     |   NVLink 7, HBM4E, Groq LP35, 800 VDC power, 1 MW racks
     |
2028 |-- FEYNMAN (roadmap) -- Feynman GPU (stacked dies), Rosa CPU,
     |   LP40 LPU, BlueField-5, CX10, copper + co-packaged optics
     v
   ==  shipping or ramping      --  roadmap only, subject to change

ASCII chart: Nvidia's generational compute claims

The chart below shows Nvidia's claimed peak low-precision compute per GPU package. These are Nvidia's figures, not independent measurements, and they are not strictly like-for-like. Blackwell figures are dense FP4. Rubin's 50 petaflops is Nvidia's "NVFP4 inference" figure, which its HGX page labels as sparse. Rubin Ultra is a roadmap projection for a four-die package.

Peak FP4-class compute per GPU package, PFLOPS (Nvidia claims)

B200 (dense FP4)        |#####                                    10
B300 (dense NVFP4)      |#######                                  15
Rubin (NVFP4 inference) |#########################                50
Rubin Ultra* (FP4)      |################################################## 100
                        +----+----+----+----+----+----+----+----+----+----+
                        0   10   20   30   40   50   60   70   80   90  100
* roadmap, 2027; 4-die package; not shipping

And at rack scale:

Rack-scale FP4-class compute, exaflops (Nvidia claims)

GB200 NVL72 (dense)       |##                                    0.72
GB300 NVL72 (dense)       |###                                   1.08
Vera Rubin NVL72 (infer.) |##########                            3.6
VR NVL144 CPX** (claimed) |######################                8.0
Rubin Ultra Kyber*        |######################################## ~15
                          +---------+---------+---------+---------+
                          0         4         8         12        16
*  roadmap, 2027     ** announced 2025, later pulled from roadmap

The direction is clear even with the caveats. Nvidia claims each new rack generation gives several times more low-precision compute, achieved partly through new silicon and partly through new number formats, larger NVLink domains, and much higher power per rack.

NVLink and NVLink Fusion: making many chips act like one

NVLink is Nvidia's high-speed link between GPUs. It is far faster than the PCIe slot a GPU plugs into on a normal PC. In a rack such as the NVL72, every GPU connects through NVLink Switch chips to every other GPU, so 72 GPUs can share memory and work as one large accelerator. Nvidia calls this scale-up networking, to distinguish it from scale-out, which links racks together over Ethernet or InfiniBand.

The per-GPU bandwidth by generation:

• NVLink 5 (Blackwell): 1.8 TB/s per GPU; 130 TB/s across an NVL72 rack.
• NVLink 6 (Rubin): 3.6 TB/s per GPU; 260 TB/s across a Vera Rubin NVL72 rack, with in-network computing for "collective" operations such as summing results across GPUs.
• NVLink 7 (Rubin Ultra): on the 2027 roadmap; no bandwidth published.

NVLink Fusion, announced in 2025, licenses NVLink technology to other companies. It lets hyperscalers and chip designers connect their own custom CPUs or accelerators (Nvidia calls these "XPUs") to Nvidia GPUs inside Nvidia's MGX rack designs. Nvidia's page names Amazon's Annapurna Labs as the first partner on a custom memory architecture called NVHBM. As part of their September 2025 deal, which included a $5 billion Nvidia investment in Intel, Intel agreed to build x86 CPUs that work with Nvidia platforms through NVLink Fusion. Samsung Foundry will offer design and manufacturing services for custom CPUs and XPUs.

The strategy matters because Nvidia's biggest customers, including Google, Amazon, Microsoft, and Meta, all design their own AI chips. NVLink Fusion is Nvidia's way of staying in the rack even when the accelerator next to it is not Nvidia's.

Networking: Spectrum-X, Quantum InfiniBand and co-packaged optics

The fastest GPU is useless if it waits on the network. When a training run spans tens of thousands of GPUs across many racks, the network becomes part of the computer.

Nvidia sells two families of scale-out networking:

• Quantum InfiniBand: a specialized, very low-latency network traditionally used in supercomputers. The current generation is Quantum-X800.
• Spectrum-X Ethernet: Nvidia's AI-tuned Ethernet, which pairs Spectrum switches with Nvidia SuperNICs. Many cloud providers prefer it because their staff already run Ethernet. Nvidia's newest generation is Spectrum-6, built on 200G SerDes (the circuits that send bits down a wire). Spectrum-XGS extends the fabric across sites hundreds of kilometers apart so that separate data centers can act as one.

At the server end, ConnectX-8 SuperNICs give 800 Gb/s per GPU in GB300 systems. ConnectX-9 in Vera Rubin is described by The Register as a 1.6 Tb/s SuperNIC. BlueField-4 combines an 800 Gb/s ConnectX-9 with a 64-core Arm CPU. It powers Nvidia's new inference context memory storage platform (also called BlueField-4 STX), which stores and reuses the "KV cache," the short-term memory a model builds while reading a long prompt.

Co-packaged optics, explained

Inside a data center, signals travel between switches over fiber-optic cables. Traditionally, each switch port has a plug-in transceiver that converts electrical signals to light. Those transceivers use a lot of power, and at hundreds of thousands of GPUs there are millions of them.

Co-packaged optics (CPO) moves the optical conversion into the same package as the switch chip, which shortens the electrical path. Nvidia announced Spectrum-X Photonics and Quantum-X Photonics CPO switches in March 2025. Nvidia's press release gave these configurations:

Nvidia claims 3.5x better power efficiency, 4x fewer lasers, and 10x better network resiliency than traditional pluggable optics. At CES 2026 it claimed 5x better power efficiency and 5x longer uptime for its Spectrum-X Ethernet Photonics switches. Exxact's GTC 2026 recap reports Nvidia describing its CPO Ethernet switch as being in full production. In May 2026, Nvidia announced plans to invest up to $3.2 billion in Corning, tied to an optical fiber supply agreement, which shows how central optics have become.

Memory: HBM3E and HBM4

HBM is a stack of DRAM chips piled vertically and wired to the GPU through thousands of tiny connections. Because the path is short and wide, it moves far more data per second than ordinary memory.

• HBM3E is used by Blackwell (192 GB per GPU) and Blackwell Ultra (288 GB). Micron and SK hynix began mass production in early 2024, and SK hynix moved to 12-high stacks later that year.
• HBM4 became a JEDEC standard in April 2025. It doubles the interface to 2,048 bits per stack, supports up to 8 Gb/s per pin (about 2 TB/s per stack), and allows stacks of 4 to 16 layers, up to 64 GB per stack. Rubin uses HBM4: 288 GB and 22 TB/s per GPU.
• HBM4E is the next step and is planned for Rubin Ultra (1 TB per package). Samsung delivered 12-layer HBM4E samples to customers in May 2026.

Why this matters to everyone else: AI demand for HBM has squeezed overall DRAM supply. Wikipedia's RTX 50 article, summarizing press coverage, notes that memory shortages pushed graphics card prices above list price from late 2025 into 2026.

Desktop AI: DGX Spark and DGX Station

Not everyone needs a data center. Nvidia now sells two "personal AI supercomputers."

DGX Spark (shipping)

First shown in early 2025 under the codename "DIGITS," DGX Spark went on sale on October 15, 2025. It is built around the GB10 Grace Blackwell superchip, which pairs a 20-core Arm CPU (10 Cortex-X925 plus 10 Cortex-A725 cores) with a Blackwell GPU.

Nvidia says one 128 GB Spark can run inference on models of up to about 200 billion parameters and fine-tune models of up to 70 billion. At GTC 2026 it said up to four Sparks can be clustered with near-linear scaling, raising the ceiling to about 700 billion parameters. Nvidia's October 28, 2026 blog said the 64 GB version was now available from partners. Nvidia's product page does not list a price; it varies by partner and configuration.

A useful caveat: 273 GB/s is a fraction of a data center GPU's memory bandwidth. Spark's selling point is capacity, meaning the ability to load big models at all, rather than raw token speed.

DGX Station (rolling out)

DGX Station is a deskside tower built on the GB300 Grace Blackwell Ultra Desktop Superchip:

Nvidia says it handles models of up to 1 trillion parameters. The first unit went to AI researcher Andrej Karpathy on March 6, 2026, and Nvidia said partner systems would ship "in the coming months" from ASUS, Dell, GIGABYTE, MSI, and Supermicro, with HP later in the year. Wikipedia, citing July 2026 reporting, gives a price range of about $100,000 to $120,000 depending on vendor and options. That is a reported figure, not an Nvidia list price.

Robotics and physical AI: Jetson Thor, Isaac GR00T, Cosmos and Omniverse

Nvidia calls robots, self-driving cars, and smart factories "physical AI." The idea is that the same AI methods behind chatbots can be trained to see, plan, and move, if they get enough data and enough onboard computing power.

Jetson Thor (shipping through partners)

Jetson is Nvidia's line of compact computers for robots and machines. Jetson Thor is the Blackwell-based generation:

Nvidia claims 7.5x the AI performance and 3.5x the energy efficiency of the previous Jetson AGX Orin. At GTC 2026 Nvidia also made IGX Thor, an industrial version with functional-safety features for settings such as factories and medical devices, generally available.

Isaac GR00T

GR00T is Nvidia's family of open "vision-language-action" models for humanoid robots. These models take camera images, a spoken or typed instruction, and the robot's current position, then output movements. Nvidia's current page highlights GR00T 1.7, a 3-billion-parameter model released on GitHub and Hugging Face with a commercial license. Nvidia says it runs on Jetson Thor at 25 Hz using NVFP4 quantization. A medical variant, GR00T-H, was trained on the Open-H dataset of surgical video. Supporting tools include Isaac Sim (simulation), Isaac Lab-Arena (testing robot policies) and Newton, an open-source GPU physics engine.

Cosmos

Cosmos is Nvidia's platform of "world foundation models." These are models trained to predict how the physical world looks and behaves, so they can generate realistic training data and help robots reason about scenes. The current generation, Cosmos 3, combines text, image, video, sound, and action in one model. It comes in three sizes: Super (64B parameters), Nano (16B, sized for post-training on a DGX Station), and Edge (4B). Cosmos models are released under the Linux Foundation's OpenMDW 1.1 license.

Omniverse

Omniverse is Nvidia's platform for physically accurate 3D simulation and "digital twins." In robotics, Omniverse scenes feed Cosmos Transfer models to make synthetic training data. In data centers, Nvidia's Omniverse DSX Blueprint and Vera Rubin DSX AI Factory reference design, announced at GTC 2026, let companies simulate an entire AI factory, including power, cooling, and layout, before building it. DSX Air runs that simulation in software.

Consumer and PC hardware: RTX 50 series and RTX Spark

GeForce RTX 50 series (shipping)

Nvidia announced the Blackwell-based RTX 50 series at CES on January 6, 2025. Launch prices and key specs:

Source: Wikipedia's GeForce RTX 50 series article, which compiles Nvidia's announcements. Street prices have often been higher because of memory shortages.

The RTX 5090 has 1,792 GB/s of memory bandwidth on a 512-bit bus. Laptop versions arrived in March 2025, and a 12 GB RTX 5070 laptop GPU appeared in April 2026.

Is there an RTX 50 "Super" refresh? As of this writing, Nvidia's RTX 50 series product page lists no Super models, and we could not confirm any launched Super cards. In October 2026, reports said the RTX 5090 was being discontinued after it disappeared from some retailers and Nvidia's own store. Nvidia denied it was discontinued.

DLSS is Nvidia's AI upscaling and frame generation. It renders games at lower resolution or fewer frames and uses AI to fill in the rest:

• DLSS 4 (2025) brought transformer-based upscaling and Multi Frame Generation, exclusive to RTX 50.
• DLSS 4.5 (January 2026) brought a second-generation transformer model for all RTX GPUs, plus Dynamic Multi Frame Generation with up to six generated frames on RTX 50.
• DLSS 5, announced at GTC 2026, adds "3D-Guided Neural Rendering" on RTX 50. It became available on September 3, 2026, starting with NBA 2K27 and GeForce NOW.

RTX Spark and the N1X: Nvidia's first PC processor (shipping, autumn 2026)

The biggest consumer story of 2026 was not a graphics card. At Computex on June 1, 2026, Jensen Huang introduced RTX Spark, a "superchip" for Windows PCs. It was long rumored under the codename N1X and developed with MediaTek and Microsoft. In these machines Nvidia silicon is the main processor, not an add-in card. Press coverage described this as a first for Windows PCs.

Reported specifications from Computex coverage:

• a 20-core Arm CPU (Nvidia describes it as a Grace CPU; the cores were designed with MediaTek) linked to a Blackwell RTX GPU with 6,144 CUDA cores over NVLink-C2C;
• up to 128 GB of unified memory, with up to about 300 GB/s of bandwidth (as reported by Digital Today);
• about 1 petaflop of FP4 AI performance;
• made on TSMC 3 nm.

Some outlets compared its graphics performance to an RTX 5070-class GPU. Nvidia had not published that comparison. Nvidia's October 28, 2026 blog says RTX Spark laptops are available now from ASUS, Dell, HP, Lenovo, Microsoft, and MSI, starting at $2,599. Nvidia says they are as thin as 14 mm and as light as three pounds, and that 128 GB configurations are on shelves. Nvidia has said it expects more than 30 laptop and 10 desktop designs.

RTX AI PC features

Nvidia markets RTX PCs as machines for running AI locally, without the cloud. Its autumn 2026 update for RTX Spark highlights small open models such as Nemotron 3.5 Bolt (one version needs as little as 4 GB of memory), the Nemotron 3.5 Transcribe speech model, and local agents such as Hermes Agent and OpenClaw. A new NVIDIA App setting lets users split GPU resources between foreground apps and background AI agents. Adobe Photoshop and Premiere have beta support for RTX Spark. Nvidia says developers can target more than 150 million RTX AI PCs and GPUs with the same CUDA code.

Software: CUDA, Dynamo and NIM

Many analysts think Nvidia's software is a stronger advantage than its silicon.

CUDA

CUDA is Nvidia's programming platform for its GPUs, launched in 2006. Almost every major AI framework is built to run on it first. Nvidia said in September 2025 that the CUDA ecosystem had more than 6 million developers and nearly 6,000 CUDA applications. On top of CUDA sit the CUDA-X libraries for specific jobs: cuDF for data frames, cuVS for vector search, cuDNN for neural networks, and others, including newer libraries aimed at chip design. Switching away from Nvidia hardware often means rewriting or re-tuning software built on these libraries. That cost is the "moat" analysts talk about.

Dynamo

Dynamo is Nvidia's open-source inference-serving framework and the successor to Triton Inference Server. It works with the popular engines vLLM, SGLang, and TensorRT-LLM. Its main features:

• Disaggregated serving: it splits prefill and decode onto different GPUs, or onto different hardware such as Vera Rubin plus LPX.
• KV-aware routing: it sends each request to a GPU that already holds the relevant cached context, which avoids recomputing it.
• NIXL: a fast library for moving cached context between GPUs, memory, and storage.
• SLO Planner: it adjusts GPU allocation to meet response-time targets.

Nvidia cites independent benchmarks showing GB300 NVL72 with Dynamo delivering up to 50x the MoE throughput of Hopper systems. Dynamo reached version 1.0 in 2026.

NIM

NIM (Nvidia Inference Microservices) packages a model, an optimized inference engine, and standard APIs into a ready-to-run container. A developer can start one with a single command and call it like a web service. Nvidia offers free hosted NIM APIs for prototyping and free self-hosting for development through its Developer Program. Production use requires an Nvidia AI Enterprise license. NIM covers Nvidia's own Nemotron models and a broad catalog of open models.

Power and cooling: why AI racks now look like industrial equipment

Perhaps the most physical change in AI hardware is power. A typical traditional data center rack draws a few kilowatts to a few tens of kilowatts. AI racks have moved far beyond that.

Table 4: Approximate rack power by generation

Sources: European server maker 2CRSI's planning guide (estimates, not Nvidia specifications), The Register, and Nvidia's 800 VDC technical blog. Nvidia did not publish an official Vera Rubin NVL72 rack power figure at launch; The Register reported only that it would be higher than the previous generation but should not double.

Liquid cooling

All Rubin systems are liquid-cooled; according to 2CRSI, there is no air-cooled Rubin. Coolant flows through cold plates attached directly to the chips. Nvidia designed Vera Rubin to accept water as warm as 45°C (113°F). Warm-water cooling matters because it lets many facilities reject heat with simple dry coolers instead of power-hungry chillers, and it makes reusing waste heat more practical. Nvidia's OCP blog notes a liquid-cooled busbar in the Vera Rubin rack, plus 20x more energy storage to smooth the sharp power swings that happen when thousands of GPUs start and stop work together.

The move to 800-volt DC

Today's racks typically take 415 or 480 V AC from the building, convert it inside the rack, and distribute 48–54 V DC to the servers. At about 200 kW, Nvidia says this approach hits physical limits. Power supplies crowd out compute, copper busbars become very heavy, and every conversion wastes energy. Nvidia's OCP blog said that in a 1 MW rack, the conventional approach could need some 200 kg of copper busbar.

Nvidia's answer is 800 V DC: convert grid power once at the edge of the facility and send high-voltage DC straight to the racks. Nvidia claims up to 5% better end-to-end efficiency, about 45% less copper than 415 V AC distribution, and up to 30% lower total cost of ownership. It plans to support 1 MW racks from 2027, with full-scale 800 V DC deployments coinciding with Kyber. For Vera Rubin NVL72, 2CRSI reports the standard rack still uses four 110 kW power shelves converting AC to 48 V DC, with 800 V DC as an optional add-on for operators preparing for Rubin Ultra. Partners named by Nvidia include Eaton, Schneider Electric, Vertiv, Delta, and Infineon, and Foxconn is building a 40 MW data center in Taiwan designed for 800 V DC.

Pricing: what has actually been reported

Nvidia rarely publishes prices for data center products. They are negotiated, bundled with networking and support, and sold mostly through server makers and clouds. Here is what we could verify:

For a sense of scale, Nvidia reported $96.2 billion of revenue for the quarter ended July 26, 2026, with $89.0 billion from its data center segment, and guided to about $108 billion for the following quarter. At GTC in March 2026, Huang said Nvidia saw roughly $1 trillion of combined Blackwell and Vera Rubin orders and opportunity through 2027. Nvidia reports its next quarter on November 17, 2026.

Export controls and China

Nvidia's most advanced chips cannot be sold to China under US export rules. Since 2022, Nvidia has made cut-down versions to stay under the limits, and both the US and Chinese governments have repeatedly moved the goalposts.

The key events, as documented in Nvidia's filings and press coverage:

• H20: a Hopper-based chip made for China after the late-2023 rules. It has 96 GB of HBM3 and 4.0 TB/s of bandwidth, with far less compute than the H100.
• April 9, 2025: the US government told Nvidia that H20 exports to China now needed a license. Nvidia took a $4.5 billion charge for excess H20 inventory and commitments, said it could not ship another $2.5 billion of H20 orders that quarter, and expected to lose about $8 billion of H20 revenue the next quarter.
• August 2025: Nvidia and AMD agreed to pay the US government 15% of revenue from certain China chip sales in exchange for export licenses. For Nvidia, this covered H20 sales.
• August–September 2025: Chinese authorities warned domestic companies against buying the H20 and told firms including ByteDance and Alibaba not to buy the RTX Pro 6000D, another China-specific chip. Nvidia reportedly told suppliers to halt H20 work. Nvidia was also developing a Blackwell-based China chip, reported as the B30A.
• Late 2025 into 2026: Washington said it would allow exports of the more capable H200 to China under conditions. Reports in early 2026 said Chinese customs blocked H200 shipments even after US clearance. In May 2026, the US reportedly cleared ten Chinese buyers, including Alibaba, Tencent, ByteDance, and JD.com, plus distributors Lenovo and Foxconn, to purchase H200s.
• August 2026: Nvidia's outlook for its current quarter assumes no data center compute revenue from China at all.

The practical result is that China, which provided $17.1 billion (13.1%) of Nvidia's revenue in its fiscal 2025, contributes little to its AI data center business today. Chinese firms are turning to domestic accelerators, while US policy keeps shifting between restriction and licensed sales.

How competitors are answering

Nvidia is dominant but not alone. Here is a brief look at the main alternatives, using each company's published figures.

• AMD Instinct MI350X / MI355X (launched June 2025): built on TSMC 3 nm with 288 GB of HBM3E at 8 TB/s, matching Blackwell Ultra's capacity. The MI355X is rated up to 9.2 petaflops of FP4 at a 1,400 W board power. AMD's next MI400 series, which includes the MI450, is among the first products to use HBM4. AMD and OpenAI announced a 6-gigawatt, multi-year deal in October 2025, with the first 1 GW of MI450 deployment set to begin in the second half of 2026. AMD's rack-scale design for this generation is called Helios. We could not verify its full specifications from primary sources in time for this article.
• Google TPU Ironwood (seventh-generation TPU, announced April 2025): 4,614 FP8 teraflops per chip, 192 GB of HBM at 7.37 TB/s, and pods of up to 9,216 liquid-cooled chips totaling 42.5 exaflops. Google claims about twice the performance per watt of its previous Trillium TPU. TPUs are available only through Google Cloud.
• Amazon Trainium3: 144 GB of HBM3E per chip at 4.9 TB/s, with UltraServers of up to 144 chips delivering 362 MXFP8 petaflops and 20.7 TB of HBM. AWS claims up to 2x the compute of Trainium2. AWS also plans to deploy more than 1 million Nvidia GPUs and Groq 3 LPUs, so it is a customer as well as a rival.
• Intel: still sells Gaudi accelerators, but its biggest AI move in this period was a partnership. Nvidia invested $5 billion in Intel in September 2025, and Intel is building x86 CPUs that connect to Nvidia GPUs over NVLink Fusion.

The pattern: hyperscalers build custom chips for their own biggest workloads and to gain bargaining power, while AMD competes most directly on memory capacity and price. Nvidia's response, through NVLink Fusion, Dynamo, and its networking, is to be the platform that everything else plugs into.

What it all means

Three trends tie Nvidia's 2026 lineup together.

First, the unit of computing is now the rack, and soon the building. Nvidia talks less about individual GPUs and more about NVL72 racks, SuperPODs, and gigawatt-scale "AI factories." Its reference designs now cover power, cooling, and the digital twin of the facility itself.

Second, inference is the main event. Blackwell Ultra's upgrades, Vera Rubin's NVFP4 inference figure, the Groq licensing deal, Dynamo, and the Vera CPU's "agentic" positioning all target the cost and speed of serving models to users, not just training them.

Third, Nvidia is pushing outward. In 2026 it entered Windows PCs as a main processor with RTX Spark, put a 748 GB workstation on desks, shipped robot computers with open models, and opened NVLink to rivals' chips. The common thread is CUDA: the more places Nvidia's software runs, the harder it is to dislodge.

The open questions are also clear. Will power grids and cooling keep up with 600 kW and then 1 MW racks? Will customers' custom chips take a meaningful share? Will China rebuild its AI stack without Nvidia? And do the generational claims hold up in independent tests once Vera Rubin is widely deployed?

FAQ

What is Nvidia's newest AI chip as of October 2026? The Rubin GPU, part of the Vera Rubin platform. It is shipping to cloud providers and large customers in Vera Rubin NVL72 racks and HGX Rubin NVL8 servers. Rubin Ultra (2027) and Feynman (2028) are announced but not shipping.

Is Vera Rubin actually available? Yes, through major clouds and large buyers. Nvidia said in August 2026 that racks were running at CoreWeave, Google Cloud, Microsoft Azure, Oracle, and Nebius, and Oracle announced general availability on October 27, 2026. Volume is still ramping.

How much faster is Rubin than Blackwell? Nvidia claims up to 5x the inference performance, up to 10x lower cost per token for MoE models, and 4x fewer GPUs for MoE training. Per GPU, Rubin is rated at 50 petaflops of NVFP4 inference against Blackwell Ultra's 15 dense, but those figures are not measured the same way. Wait for independent benchmarks such as MLPerf before treating any multiple as typical.

What happened to "Vera Rubin NVL144"? It is mostly a naming change. Nvidia briefly planned to count each GPU die separately, which would have doubled the number in the name. It reverted to counting packages, so the rack is now called NVL72.

What is the difference between a GPU and an LPU? A GPU is a flexible parallel processor with large external HBM memory, good at both training and inference. Groq's LPU, which Nvidia now uses under license in the Groq 3 LPX, keeps a smaller amount of very fast SRAM on the chip and runs a fixed, pre-planned schedule. That suits the token-by-token decode step of inference.

Did Nvidia buy Groq? Not outright. In December 2025, Groq signed a non-exclusive licensing agreement with Nvidia, and Groq's founder, its president, and other staff joined Nvidia. CNBC put the value at about $20 billion, which neither company confirmed. Groq continues as an independent company.

Is Rubin CPX cancelled? Nvidia removed it from its GTC 2026 roadmap without a full explanation. An analyst reported in September 2026 that a redesigned HBM4 version is planned for 2027. Nvidia has not confirmed that report.

Can I buy a DGX Spark or DGX Station? DGX Spark has been on sale since October 2025 through Nvidia and partners such as Acer, ASUS, Dell, GIGABYTE, HP, Lenovo, and MSI. A 64 GB version is now available through partners. DGX Station systems are rolling out from partners in 2026, at a reported $100,000–$120,000.

What is RTX Spark, and is it the same as N1X? RTX Spark is the product name for Nvidia's Arm-based PC superchip, which was known by the codename N1X before launch. It combines a 20-core CPU and a 6,144-core Blackwell GPU with up to 128 GB of unified memory. Laptops started at $2,599 in autumn 2026.

Will there be an RTX 50 Super series? We could not confirm any announced or released RTX 50 Super cards as of October 31, 2026. Nvidia's product page lists none.

Why do AI data centers need liquid cooling and 800-volt power? Because racks now draw from about 135 kW (GB300) to an estimated 190–230 kW (Vera Rubin), and Nvidia is planning for 600 kW (Kyber) and eventually 1 MW. Air cannot carry away that much heat, and low-voltage power distribution needs too much copper and wastes too much energy at those levels.

Can Nvidia sell its AI chips in China? Only in restricted forms, and with changing rules on both sides. The H20 required a US license from April 2025, and Chinese authorities have discouraged purchases of some Nvidia chips. The US has cleared some H200 sales to approved buyers. Nvidia's current outlook assumes no data center compute revenue from China.

Who are Nvidia's main competitors in AI chips? AMD (the Instinct MI350 and the upcoming MI400/MI450 with the Helios rack), Google (the TPU Ironwood), Amazon (Trainium), and to a lesser degree Intel. Several of these companies are also major Nvidia customers or partners.

What is NVFP4, and why does everyone quote it? NVFP4 is Nvidia's 4-bit number format. Using fewer bits per number lets chips do far more math per second and fit bigger models in memory, with accuracy Nvidia says stays close to 8-bit. Headline "petaflops" figures for Blackwell and Rubin are usually NVFP4, so they cannot be compared directly with older FP16 or FP8 figures.

Sources and further reading

• NVIDIA Blog, GTC 2026 news roundup: https://blogs.nvidia.com/blog/gtc-2026-news/
• NVIDIA Newsroom, Rubin platform press release (CES 2026): https://nvidianews.nvidia.com/news/rubin-platform-ai-supercomputer
• The Register, "Every conference is an AI conference as Nvidia unpacks its Vera Rubin CPUs and GPUs at CES": https://www.theregister.com/2026/01/05/ces_rubin_nvidia/
• NVIDIA, GB300 NVL72 product page: https://www.nvidia.com/en-us/data-center/gb300-nvl72/
• NVIDIA, GB200 NVL72 product page: https://www.nvidia.com/en-us/data-center/gb200-nvl72/
• NVIDIA, HGX platform page (HGX B200, B300, Rubin NVL8): https://www.nvidia.com/en-us/data-center/hgx/
• NVIDIA Developer Blog, "Inside NVIDIA Blackwell Ultra": https://developer.nvidia.com/blog/inside-nvidia-blackwell-ultra-the-chip-powering-the-ai-factory-era/
• NVIDIA Newsroom, Rubin CPX announcement: https://nvidianews.nvidia.com/news/nvidia-unveils-rubin-cpx-a-new-class-of-gpu-designed-for-massive-context-inference
• Wccftech, Rubin CPX off the roadmap: https://wccftech.com/nvidia-rubin-cpx-is-off-the-roadmap-but-may-return-with-feynman-in-2028/
• TechNews (Taiwan), Rubin CPX re-enters roadmap (Kuo report): https://technews.tw/2026/09/02/nvidia-rubin-cpx-reenters-roadmap-hbm4-ai-prefill/
• StorageReview, "NVIDIA Groq 3 LPX: everything we know": https://www.storagereview.com/news/nvidia-groq-3-lpx-everything-we-know
• Exxact, GTC 2026 recap: https://www.exxactcorp.com/blog/news/nvidia-gtc-2026-recap-era-of-tokens-and-inference
• NVIDIA Blog, gigawatt AI factories, OCP and Vera Rubin: https://blogs.nvidia.com/blog/gigawatt-ai-factories-ocp-vera-rubin
• NVIDIA Developer Blog, 800 V HVDC architecture: https://developer.nvidia.com/blog/nvidia-800-v-hvdc-architecture-will-power-the-next-generation-of-ai-factories/
• 2CRSI, Blackwell to Rubin power and cooling guide: https://2crsi.com/nvidia-from-blackwell-to-rubin-platforms-power-and-cooling
• NVIDIA, NVLink Fusion: https://www.nvidia.com/en-us/data-center/nvlink-fusion/
• NVIDIA Newsroom, Spectrum-X and Quantum-X co-packaged optics switches: https://nvidianews.nvidia.com/news/nvidia-spectrum-x-co-packaged-optics-networking-switches-ai-factories
• NVIDIA Newsroom, NVIDIA and Oracle expand Vera Rubin (Oct 27, 2026): https://nvidianews.nvidia.com/news/nvidia-and-oracle-expand-nvidia-vera-rubin-to-multi-gigawatt-scale-and-bring-nvidia-vera-cpu-to-oracle-cloud-infrastructure
• NVIDIA Newsroom home: https://nvidianews.nvidia.com/
• NVIDIA Newsroom, Q2 fiscal 2027 financial results: https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027
• NVIDIA Newsroom, Q1 fiscal 2026 financial results (H20 charge): https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-first-quarter-fiscal-2026
• NVIDIA Newsroom, Q3 fiscal 2027 conference call notice: https://nvidianews.nvidia.com/news/nvidia-sets-conference-call-for-third-quarter-financial-results-6928873
• NVIDIA, DGX Spark product page: https://www.nvidia.com/en-us/products/workstations/dgx-spark/
• NVIDIA Newsroom, DGX Spark arrives: https://nvidianews.nvidia.com/news/nvidia-dgx-spark-arrives-for-worlds-ai-developers
• NVIDIA, DGX Station product page: https://www.nvidia.com/en-us/products/workstations/dgx-station/
• NVIDIA, Jetson Thor: https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-thor/
• NVIDIA, Isaac GR00T: https://www.nvidia.com/en-us/industries/robotics/isaac-groot/
• NVIDIA, Cosmos: https://www.nvidia.com/en-us/ai/cosmos/
• NVIDIA, NIM microservices: https://www.nvidia.com/en-us/ai-data-science/products/nim-microservices/
• NVIDIA Developer, Dynamo: https://developer.nvidia.com/dynamo
• NVIDIA, GeForce RTX 50 series: https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/
• NVIDIA Blog, RTX Spark laptops update (Oct 28, 2026): https://blogs.nvidia.com/blog/local-ai-rtx-spark-laptops-updates-models-agents-games/
• Free Press Journal, N1X unveiled at Computex 2026: https://www.freepressjournal.in/tech/nvidia-unveils-n1x-chip-for-windows-powered-laptops-at-computex-2026-heres-why-it-is-a-big-deal
• Digital Today, Nvidia enters PC market with N1X: https://www.digitaltoday.co.kr/en/view/59955/nvidia-declares-entry-into-pc-market-unveils-arm-based-n1x-chip
• Wikipedia, GeForce RTX 50 series: https://en.wikipedia.org/wiki/GeForce_RTX_50_series
• Wikipedia, Rubin (microarchitecture): https://en.wikipedia.org/wiki/Rubin_(microarchitecture)
• Wikipedia, Nvidia: https://en.wikipedia.org/wiki/Nvidia
• Wikipedia, Nvidia DGX: https://en.wikipedia.org/wiki/Nvidia_DGX
• Wikipedia, Hopper (microarchitecture): https://en.wikipedia.org/wiki/Hopper_(microarchitecture)
• Wikipedia, High Bandwidth Memory: https://en.wikipedia.org/wiki/High_Bandwidth_Memory
• Wikipedia, AMD Instinct: https://en.wikipedia.org/wiki/AMD_Instinct
• AMD, AMD and OpenAI 6-gigawatt partnership: https://ir.amd.com/news-events/press-releases/detail/1260/amd-and-openai-announce-strategic-partnership-to-deploy-6-gigawatts-of-amd-gpus
• Google, Ironwood TPU announcement: https://blog.google/products/google-cloud/ironwood-tpu-age-of-inference/
• AWS, Trainium: https://aws.amazon.com/ai/machine-learning/trainium/