Nvidia is making the biggest bet in AI hardware history and most people scrolled past it. A $20 billion deal signed on Christmas Eve. A two-tonne supercomputer that assembles in five minutes. A chip that makes AI ten times cheaper to run. This is not another product launch. This is Nvidia quietly rebuilding the entire foundation that powers every AI tool you use daily. From the Vera Rubin NVL72 to the Groq acquisition to a $30 billion OpenAI investment, three moves just reshaped the future of artificial intelligence. Here is everything that happened and exactly what it means for you.
But if you sit with it, if you trace the logic from one decision to the next, what Nvidia has done across the last few months is not just launch new products. They have redesigned what it means to build and run artificial intelligence at scale. And the consequences are going to ripple through every product you use.
The Machine Most People Have Never Heard Of
Start with the hardware.
Nvidia’s Vera Rubin NVL72 is a rack-scale AI supercomputer. Not a chip. A rack. A full, assembled, liquid-cooled tower of compute that sits in data centers and runs the models behind the AI tools millions of people use every day.
At the rack level, the flagship configuration is 72 Rubin GPUs and 36 Vera CPUs connected through NVLink 6. The Rubin GPU is built on TSMC’s 3nm process with 336 billion transistors across two reticle-limited compute dies and two I/O dies. Each GPU package carries 288GB of HBM4 memory delivering approximately 22TB per second of bandwidth, a 2.8x improvement over Blackwell’s 8TB per second.
That last number is worth pausing on. Nearly three times the memory bandwidth of the previous generation. If the previous generation was a highway with four lanes, Rubin is a twelve-lane expressway, and cars are moving faster on every one of them.
At the full rack level, Nvidia’s own figures cite 3.6 exaflops of NVFP4 inference and 2.5 exaflops of training, plus 20.7TB of HBM4 capacity and 54TB of LPDDR5x capacity, along with 1.6 petabytes per second of HBM bandwidth.
An exaflop is one quintillion floating-point operations per second. The Vera Rubin NVL72 does 3.6 of those, just for inference. That is a number that loses meaning when written out in full. The simpler way to understand it is this: the things you ask AI to do, answering questions, writing code, analyzing documents, reasoning through complex problems, Vera Rubin can do them faster, cheaper, and in larger volume than anything that came before.
Nvidia claims Rubin delivers one-tenth the cost per million tokens compared to Nvidia Blackwell for highly interactive, deep reasoning agentic AI, and trains mixture-of-expert models with one-fourth the number of GPUs over the Blackwell architecture.
One-tenth the cost. That is not a marginal improvement. That is a restructuring of the economics of AI deployment.
Why Moore’s Law Broke and Nvidia Had to Invent a New Strategy
There is a reason Nvidia had to build an entire rack instead of just building a better chip. Jensen Huang explained it himself during his CES keynote.
Moore’s Law has largely slowed, and the number of transistors that can be added year after year can no longer keep up with the demand for ten times larger AI models.
For decades, the chip industry ran on a reliable rhythm. Every two years, chips doubled in power without doubling in cost. Engineers could sit back and let physics do the hard work. Design a chip, ship it, and wait for the next generation to be faster.
That rhythm has broken down. Rubin features 1.6 times more transistors than Blackwell, which is the starting point for the performance boost. But 1.6 does not get you to 10 times. “It is impossible to keep up with those kinds of rates,” Huang said during his keynote, “unless we deployed aggressive, extreme co-design — basically innovating across” the entire system.
That word, co-design, is the key to understanding everything Nvidia has done here. Rubin is the result of what the company calls extreme co-design across six types of chips — the Vera CPU, the Rubin GPU, the NVLink 6 switch, the ConnectX-9 SuperNIC, the BlueField-4 data processing unit, and the Spectrum-6 Ethernet switch. Those building blocks all come together to create the Vera Rubin NVL72 rack.
Six chips. Each one designed to work with the others. No single chip becomes a bottleneck because every chip was designed with every other chip in mind from the beginning. This is categorically different from how hardware is usually built, where chips are designed independently and then forced to cooperate through software abstractions and workarounds.
When Nvidia says it treats the data center as the unit of compute, not the chip, this is what they mean. The rack is the product. The individual chips are components inside that product.
Assembly That Takes Five Minutes, And Why That Changes Everything
Here is a detail that nearly every headline skipped.
Although the compute tray must be taken offline during maintenance, the modular cable-free design reduces service time by up to 18 times. Assembly that used to take more than 1.5 hours for Blackwell now takes only about five minutes with Vera Rubin.
Five minutes.
The previous generation of hardware took a trained technician over ninety minutes to assemble a single rack. Vera Rubin takes five. The entire system is fanless, tubeless, and cableless. Everything connects through modular trays that slide in and out. The entire system is 100 percent cooled with liquid cooling.
Why does assembly time matter? When you are deploying hundreds of thousands of these racks across multiple data centers, the difference between ninety minutes and five minutes per rack is the difference between weeks and days of deployment time. Faster deployment means the hardware is generating revenue sooner. It means companies can scale capacity in response to demand instead of planning months in advance. It means when a component fails at 2am, a technician can replace it and have the system back online before sunrise instead of taking an entire cluster offline for a day.
The rack’s modular, cable-free tray design enables up to 18 times faster assembly and servicing than Blackwell. The number sounds like marketing until you think about the logistics of operating a data center at scale. Then it sounds like the engineers finally got frustrated and fixed something that should have been fixed years ago.
The Inference Revolution Nobody Planned For
Here is the part of this story that requires a little background, because without it the Groq deal does not make sense.
Training an AI model is the part everyone talks about. You feed a model billions of documents over weeks or months on thousands of GPUs, and it learns patterns. This is computationally expensive and happens relatively infrequently, once per major model version.
Inference is the other part. It is what happens every single time you ask ChatGPT a question, every time a coding assistant suggests the next line, every time a customer service bot reads your complaint and formulates a response. Inference is not the one-time cost. It is the ongoing cost. It is running, constantly, at enormous scale, for millions of users simultaneously.
For years, the AI industry assumed that the same GPUs used for training could handle inference just fine. Train once, deploy forever, cost of inference is manageable. That assumption held when AI was doing simple things. It broke when AI started doing complex things.
The word-by-word generation process currently plagues large-scale AI agents. The shift is critical as agentic AI — autonomous systems that perform tasks — becomes the primary driver of enterprise tech spending in 2026.
Agentic AI is the new frontier. Instead of a single question and a single answer, AI systems are now asked to plan multi-step tasks, call tools, remember context across long conversations, coordinate with other AI agents, and revise their thinking when earlier steps produce unexpected results. Each of those steps requires inference. A single complex agentic task might require dozens or hundreds of inference calls, each one generating tokens, each one drawing on the same overloaded GPU infrastructure.
The bottleneck is no longer training. The bottleneck is inference speed and cost. And Nvidia, having built its dominance on training GPUs, suddenly found itself in a race to catch up in a market it helped create.
The $20 Billion Move That Rewrote the Rules
On Christmas Eve 2025, while most of the technology industry was on holiday, Nvidia quietly disclosed it had signed a licensing agreement with Groq.
The price tag was approximately $20 billion. The structure was unusual. The acquisition is structured as a licensing-and-acquihire agreement, effectively transferring all of Groq’s key assets to Nvidia while allowing Groq to remain a nominally independent company. This arrangement lets Nvidia circumvent rigorous antitrust scrutiny by maintaining the appearance of competition, even as it absorbs Groq’s intellectual property, key engineers including founder Jonathan Ross and President Sunny Madra, and architectural innovations.
To understand why Nvidia paid $20 billion for a company that most people outside the AI hardware world had never heard of, you need to understand what Groq actually built.
Groq, founded in 2016 by ex-Google TPU lead Jonathan Ross, built specialized Language Processing Units for AI inference. Its chips emphasize a deterministic, single-core design with massive on-chip SRAM, delivering remarkably low-latency inference performance that in independent tests ran roughly two times faster than any other provider’s solution. This is in stark contrast to Nvidia’s GPUs, which evolved from graphics processors and rely on many cores plus off-chip HBM memory, introducing overhead and variability.
A Language Processing Unit, or LPU, is not a general-purpose chip. It does one thing — run AI models to generate outputs — and it does that one thing with a kind of focused ferocity that general-purpose hardware cannot match. Groq’s architecture achieves up to tens of terabytes per second of memory bandwidth via on-chip SRAM and avoids wasted cycles through its static scheduling and compiler-driven execution.
Think of it this way. A GPU is like a highly trained generalist who can do almost anything but takes a moment to figure out the best approach for each new task. An LPU is like a specialist who has done one specific job ten thousand times and can execute it from muscle memory before you finish explaining what you need.
Why OpenAI Was Shopping Around
Before Nvidia stepped in, OpenAI had begun conversations with two chip companies — Cerebras and Groq — in search of better inference processing solutions. Those talks were cut short when Nvidia finalized the $20 billion licensing agreement with Groq, effectively removing OpenAI’s ability to partner directly with the startup.
This detail matters enormously. OpenAI, the company behind ChatGPT, the largest consumer AI product in history, was looking for alternatives to Nvidia. Not because Nvidia’s hardware is bad. Because inference at scale has specific requirements that general-purpose training GPUs were not optimized to meet. OpenAI wanted faster, cheaper inference. It was willing to go outside its primary hardware supplier to get it.
Nvidia’s response was to spend $20 billion to ensure that when the best inference chip startup in the world came to market, it came to market through Nvidia.
The business relationship between Nvidia and OpenAI goes far beyond simple hardware transactions. Last September, Nvidia revealed plans to invest up to $100 billion in OpenAI. This deal granted Nvidia equity ownership in the AI company while providing OpenAI with capital to purchase advanced processors. Nvidia now functions as both supplier and investor — a dual position that creates strong incentives to control OpenAI’s hardware procurement decisions.
Supplier and investor simultaneously. Think about that for a moment. Nvidia is selling hardware to OpenAI. Nvidia also owns a stake in OpenAI. When OpenAI succeeds, Nvidia profits twice. And when OpenAI needs hardware, Nvidia is the most motivated seller in the room.
OpenAI announced a massive purchase of dedicated inference capacity from Nvidia, bolstered by a $30 billion investment from the chip giant.
The inference chip announcement, the Groq deal, and the OpenAI investment are not three separate stories. They are one story. Nvidia identified the next battlefield, bought the best weapon on that battlefield, and locked in the most important customer before the battle officially started.
The Startups Left in the Wake
The Groq deal sent ripples through the AI chip startup ecosystem immediately.
The deal bolsters the standing of other startups building their own AI chips, including Cerebras, D-Matrix, and SambaNova. It also lifts AI inference software platform startups like Etched, Fireworks, and Baseten, strengthening their valuations and making them more attractive acquisition targets in 2026, according to analysts, founders, and investors.
When the largest company in the AI hardware space pays $20 billion for an inference chip startup, it sends a signal to every investor and every acquirer in the market: inference-focused chip companies are worth serious money.
Sid Sheth, CEO of D-Matrix, told Fortune: “When the Nvidia-Groq deal happened, we said, ‘Finally, the market recognizes it.’”
Cerebras CEO Andrew Feldman was more pointed. In the past, the perception that Nvidia GPUs were all you needed for AI acted as a moat, keeping AI chip startups from gaining ground. But that moat is now gone with the Groq deal. “It reflects a growing industry reality — the inference market is fragmenting, and a new category has emerged where speed is not a feature — it is the entire value proposition.”
The inference market is fragmenting. That is a phrase worth sitting with. What it means in practice is that the age of one chip doing everything is ending. Training will have its chips. Inference will have its chips. Edge deployment will have its chips. The hardware stack underneath AI is becoming specialized in the same way that AI applications themselves are becoming specialized.
Nvidia saw this coming. The Groq deal was not a panic buy. It was the completion of a strategic picture that Nvidia had been assembling for at least two years.
The Companies Already Signed Up
The list of companies committed to deploying Vera Rubin is a directory of the most consequential names in modern computing.
Among the first cloud providers to deploy Vera Rubin-based instances in 2026 will be AWS, Google Cloud, Microsoft, and Oracle Cloud Infrastructure, as well as Nvidia Cloud Partners CoreWeave, Lambda, Nebius, and Nscale.
Microsoft will deploy Nvidia Vera Rubin NVL72 rack-scale systems as part of next-generation AI data centers, including future Fairwater AI superfactory sites. Microsoft’s next-generation Fairwater AI superfactories featuring Nvidia Vera Rubin NVL72 rack-scale systems will scale to hundreds of thousands of Nvidia Vera Rubin superchips.
Hundreds of thousands of superchips. In a single factory. That number is almost too large to hold. Each superchip is already a formidable piece of hardware. Hundreds of thousands of them, networked together, represents a concentration of AI compute that would have seemed like science fiction five years ago.
Every AI product you use that runs on AWS, Azure, or Google Cloud infrastructure will eventually run on this hardware. The migration will not happen overnight. But the contracts are signed and the roadmap is clear.
What Competition Looks Like Now
Nvidia will tell you with confidence that they are not worried. And they have earned the right to that confidence. They hold approximately 90 percent of the AI chip market. Their software ecosystem, CUDA, is embedded in virtually every AI research workflow in the world. Switching away from Nvidia is not just buying different hardware. It is rewriting years of optimized code.
But the competition is real.
AMD’s Helios rack systems promise to deliver floating point performance roughly equivalent to Nvidia’s Vera Rubin NVL72 at 2.9 exaflops versus 2.5 to 3.6 exaflops of FP4. For applications that cannot take advantage of Nvidia’s adaptive compression technology, Helios is, at least on paper, faster. AMD still has a capacity lead with 432GB of HBM4 per GPU socket compared to 288GB on Rubin. In theory, this should allow AMD-based systems to serve 50 percent larger MoE models on a single double-wide rack.
On paper, AMD is closer than most people realize. In practice, the software ecosystem gap remains enormous. Nvidia’s CUDA libraries, its networking software, its inference frameworks, its relationships with cloud providers, all of it creates friction that pure hardware performance cannot easily overcome.
Broadcom is building custom AI chips for Google and Meta. Google’s TPUs have been running inference for years. China is developing domestic alternatives, with some reports claiming 14nm chips that rival earlier Nvidia generations.
The moat Nvidia has built is not hardware. It is the full stack — hardware, networking, software, relationships — all co-designed to work together. Matching one piece of that stack is achievable. Matching all of it simultaneously is genuinely difficult.
The Energy Question That Has No Easy Answer
There is an honest complication buried in the Vera Rubin story and it deserves acknowledgment.
All of this raw compute and bandwidth is impressive on its face, but the total cost of ownership picture is likely most important to Nvidia’s partners as they ponder massive investments in future capacity.
More performance means more power. The Vera Rubin rack requires substantially more electricity than Blackwell. Nvidia’s defense is efficiency — ten times more performance per watt — but that defense only holds at the level of individual computations. When you scale to Microsoft’s Fairwater superfactories with hundreds of thousands of chips, total electricity consumption rises dramatically regardless of per-chip efficiency improvements.
The entire system moves to liquid cooling, which reduces water consumption compared to evaporative cooling systems. But it increases infrastructure complexity for data centers that were not designed to route liquid coolant through their floors.
There is a broader pattern here that no single company can solve. When computing becomes cheaper, usage expands. More efficient AI means more AI models, more inference queries, more applications, more users. Total energy demand tends to rise even as efficiency per computation improves. The industry is running faster toward a wall that efficiency improvements keep pushing further away but cannot eliminate entirely.
Nvidia is setting the pace of that run. That comes with responsibility that quarterly earnings calls do not fully capture.
Named After Someone Who Found Invisible Things
Nvidia named this platform after Vera Florence Cooper Rubin, an American astronomer who spent her career studying galaxies and their rotation patterns.
What she found was perplexing. Galaxies were rotating in ways that should have torn them apart, based on the visible mass they contained. Stars at the outer edges of galaxies were moving too fast. The gravitational math did not work.
Rubin’s conclusion was that there was more mass in these galaxies than anyone could see. Mass that did not emit or reflect light. Mass that could only be detected by its gravitational effect on everything around it. Her work provided some of the first solid observational evidence for what we now call dark matter, a substance that makes up roughly 27 percent of the universe’s mass-energy content and remains invisible to every instrument we have built.
She spent years finding things that could not be seen directly, only inferred from their effects on everything else.
There is something fitting in that parallel. You will never see a Vera Rubin NVL72 rack. You will never touch a Rubin GPU or watch a Groq LPU generate a token. But you will feel the effects of this hardware in every AI interaction you have from 2026 onward. Faster responses. Lower costs. More capable reasoning. AI agents that can maintain context across genuinely complex tasks without degrading.
The infrastructure is invisible. Its effects are everywhere.
What This Actually Means for You
If you use any AI product regularly — a writing assistant, a coding tool, a search engine with AI features, a customer service system — the hardware your queries run on is about to change significantly.
Wall Street is predicting a 41 percent surge in Nvidia shares in 2026, compared with a 34.8 percent gain in 2025. Markets are betting that the Vera Rubin platform and the Groq-powered inference chip represent Nvidia’s next phase of dominance, not just a product cycle.
But the investor story is secondary to the user story. The real consequence of ten times lower inference cost is not a higher stock price. It is that AI tools become affordable for a dramatically wider range of companies. Small businesses that could not justify the expense of running AI internally will be able to. Industries that have been waiting for AI to be cheap enough to deploy at scale, healthcare, education, logistics, will find the math suddenly working in their favor.
The move reflects a broader industry shift away from pure AI training toward continuous, 24/7 inference computing, positioning Nvidia at the center of the next phase of AI adoption.
The next phase of AI adoption is not a new model from OpenAI or a flashier chatbot interface. It is the quiet, structural work of making AI responses fast enough, cheap enough, and reliable enough to run inside every application, constantly, invisibly, the way databases and networks run today.
Vera Rubin is Nvidia’s answer to how that infrastructure gets built. The Groq deal is Nvidia’s answer to who controls the most critical component of that infrastructure. The OpenAI partnership is Nvidia’s answer to who the first and most important customer of that infrastructure will be.
Three moves. One strategy. And a hardware platform named after a woman who spent her life proving that the most important things are often the ones nobody can see.