The hype around Nvidia’s Blackwell architecture has reached a fever pitch, but if you’re only looking at the massive headlines about Meta’s 1.3 million unit target, you’re missing the actual story. Honestly, the real drama isn't just with Mark Zuckerberg's social media empire. It is in the desperate, multi-billion dollar scramble by everyone else—the "neoclouds," the old-school tech giants, and the sovereign nations trying to build their own AI "factories" before they get left in the dust.
Right now, in early 2026, the situation is basically a high-stakes game of musical chairs where the chairs cost $40,000 each and the music never stops. While Meta is a huge chunk of the pie, the nvidia blackwell gpu orders excluding meta tell a much more interesting tale of industrial shift. We are talking about over 3.6 million units claimed by other players. Microsoft, Amazon, Google, and Oracle are essentially vacuuming up every piece of silicon TSMC can package.
The 3.6 Million Unit Reality
Jensen Huang recently confirmed that Blackwell orders—once you strip Meta out of the equation—totaled a staggering 3.6 million units for the initial ramp. To put that in perspective, previous flagship launches like the H100 Hopper never even dreamed of hitting seven-figure volumes in their first few quarters.
Why the sudden explosion? Because the world has moved from "let's see if this AI thing works" to "we need to train 100-trillion parameter models or our stock price dies."
💡 You might also like: ChatGPT Search: What Most People Get Wrong One Year Later
Microsoft and the OpenAI Hunger
Microsoft is arguably the hungriest player here. They aren't just buying for Azure; they are buying to keep OpenAI’s "Stargate" project on life support. You've probably heard the rumors of the $100 billion to $500 billion supercomputer. Well, that machine doesn't run on hopes and dreams. It runs on the GB300 NVL72 racks.
Microsoft’s commitment is so massive that they’ve reportedly had to redesign entire data centers just to handle the liquid cooling requirements. Air cooling? Forget about it. These Blackwell racks pull so much power—about 120kW per rack—that you basically need a dedicated river to keep them from melting.
Amazon’s "Project Ceiba" Twist
Amazon is doing something a bit different. They’re buying Blackwell, sure, but they’re also trying to hedge their bets. Their "Project Ceiba" is a massive supercomputer co-built with Nvidia that uses 20,736 GB200 Grace Blackwell Superchips.
But here’s the kicker: Amazon is also pushing its own Trainium chips. They are the only ones really trying to tell Nvidia, "Hey, we love you, but we don't want to be your hostage forever." Despite that, their Blackwell orders remain at record highs because, quite frankly, Trainium isn't ready to handle the heaviest Llama-4 or GPT-5 level training yet.
The Neocloud Revolution: CoreWeave and Nscale
If you want to see where the real "nvidia blackwell gpu orders excluding meta" are going, look at the companies you hadn't heard of three years ago.
👉 See also: NASA Parallel Universes 2024: What Actually Happened With That Viral Antarctic Story
CoreWeave is the poster child for this. They started as a crypto mining outfit and now they have a $55.6 billion revenue backlog. Think about that for a second. That is more than the GDP of some countries, all sitting in a queue for GPU rentals. In late 2025, Nvidia even inked a deal to guarantee $6.3 billion in compute capacity for CoreWeave. It’s a circular economy: Nvidia invests in CoreWeave, and CoreWeave uses that money to buy more Blackwell chips.
Then there’s Nscale in the UK. They just secured 120,000 Blackwell GPUs to build Europe’s largest AI cluster. This is part of the "Sovereign AI" movement. Governments are realizing that if they don't own the compute, they don't own their future.
- Oracle: Scaling to 131,072 GPUs in a single zettascale cluster.
- Google: Integrating GB300s into their A4X Max VMs.
- CoreWeave: Dominating the "rent-a-GPU" market for startups.
The Supply Chain Bottleneck (It’s Not the Wafers)
You might think Nvidia's biggest problem is making the chips. It’s not. TSMC can crank out wafers all day long. The real nightmare is CoWoS-L packaging and HBM3e memory.
Blackwell is a "chiplet" design. It’s basically two giant dies glued together with high-speed interconnects. That "glue" is the CoWoS (Chip on Wafer on Substrate) technology. In early 2026, the lines for this packaging are booked out for over 12 months. Every time a memory provider like SK Hynix or Micron has a minor yield issue with their HBM3e (High Bandwidth Memory), the entire Blackwell production line grinds to a halt.
This is why Lead times for a Blackwell rack are currently sitting at 52 weeks. If you didn't order your GPUs a year ago, you aren't getting them today. Period.
📖 Related: Why the Apple Store World Trade Center is Still New York’s Most Interesting Retail Space
The "Shoulda Waited" Problem: Enter Rubin
Here is the part that is keeping CTOs up at night. Nvidia just announced the Vera Rubin platform at CES 2026.
Rubin is promised to be 5x faster for inference and 10x cheaper per token. It uses HBM4 memory and is scheduled for late 2026. This creates a bizarre paradox. Companies are spending $3 million per Blackwell rack today, knowing that in six months, Nvidia will ship something that makes their brand-new hardware look like a calculator.
Does that stop the orders? No. Because in the AI race, being six months late is the same as being dead. If Microsoft waits for Rubin while Google scales on Blackwell, Google wins the next two years of search and assistant dominance. It’s a treadmill that nobody can get off.
What This Means for Your Strategy
If you're an enterprise buyer or an investor looking at the nvidia blackwell gpu orders excluding meta, you need to look past the raw numbers. The "Blackwell wave" is less about the hardware and more about the infrastructure transition.
- Liquid Cooling is Non-Negotiable: If you are building on-prem, stop thinking about fans. You need to invest in liquid-to-chip cooling now.
- The Cloud is Safer (For Now): Because of the "Rubin Obsolescence," renting from AWS or Oracle makes more sense than buying. Let the hyperscalers deal with the hardware being "outdated" in 12 months.
- Watch the "Neoclouds": Companies like CoreWeave often get "priority" shipping because Nvidia has an equity stake in them. If you can't get Blackwell through Azure, you might find it there.
The reality of the 2026 Blackwell market is that demand still outstrips supply by a massive margin. Even without Meta's billion-dollar checks, the rest of the world is more than happy to pick up the slack. The backlog is real, the power requirements are terrifying, and the race to the next generation is already making today's "state of the art" feel like history.
To navigate this, focus on building your software stack to be hardware-agnostic. Use NVIDIA NIM microservices to ensure your models can migrate from Blackwell to Rubin without a complete rewrite of your architecture. Secure your power contracts first, because in 2026, electricity is the new oil, and the GPU is just the engine.