What Grace Blackwell NVL72 is — a rack that behaves like one GPU
Grace Blackwell NVL72 is not a chip but a rack. Nvidia's GB200 NVL72 puts 36 Grace CPUs and 72 Blackwell GPUs in one liquid-cooled cabinet and connects them through a two-tier NVLink switch fabric so software can address the whole rack as a single accelerator. The building block is the GB200 Superchip — one Grace CPU joined to two Blackwell GPUs over a 10 TB/s NVLink-C2C link — and 36 of those make a rack. Published figures give about 1.4 exaflops of AI compute per rack, roughly 130 TB/s of aggregate NVLink bandwidth within one compute domain, and around 120 kilowatts of power draw, which is why the cabinet is liquid-cooled. The reason this unit exists is memory, not arithmetic: a frontier model no longer fits in the memory of one GPU, so the bottleneck moves from computation to the wire between GPUs, and NVL72 is a product aimed at that wire. The name entered wider circulation on September 7, 2026, when Jensen Huang said GPT-6 Astra had been trained on more than 100,000 of them
The three lines
- Definition — a rack, not a chip: 36 Grace CPUs and 72 Blackwell GPUs bound by a two-tier NVLink fabric and addressed as one accelerator
- Numbers — about 1.4 exaflops per rack, roughly 130 TB/s aggregate NVLink bandwidth, around 120 kW draw, hence liquid cooling
- Reason — the bottleneck is the wire between GPUs, not the arithmetic, once a model exceeds one GPU's memory
Key questions
- What is Grace Blackwell NVL72
- **It is the name of a rack, not a chip.** Break the name into three parts and it reads plainly. **Grace** = Nvidia's **CPU**. **Blackwell** = the GPU of that generation. **NVL72** = **72 GPUs bound by NVLink**. So: **"a system that ties Grace CPUs and Blackwell GPUs together in units of 72 over NVLink."** The construction stacks up. **① GB200 Superchip** — one Grace CPU plus **two** Blackwell GPUs, joined by a **10 TB/s NVLink-C2C** link. **② The rack** — **36** of those superchips in one cabinet, which gives **36 CPUs and 72 GPUs**. **③ The fabric** — the 36 superchips are wired through a **two-tier NVLink switch**. **That third line is the point.** Putting many GPUs in one cabinet is not new. **What distinguishes NVL72 is that the 72 GPUs present to software as a single accelerator.** Rather than managing 72 cards and hand-splitting data between them, a developer treats the rack as one large memory space. **The practical consequence is that the unit of sale moved from the card to the rack.**
- What do the numbers look like
- **One rack is equivalent to a small data-center bay.** | Item | Value | |---|---| | Grace CPUs | **36** | | Blackwell GPUs | **72** | | AI compute | about **1.4 exaflops** | | Aggregate NVLink bandwidth | about **130 TB/s** (single compute domain) | | Power draw | about **120 kW** | | Cooling | **liquid** | **The three numbers explain one another.** ① **1.4 exaflops** means 1.4×10¹⁸ operations per second, but at low precision — it cannot be compared directly with supercomputer ranking figures. ② **130 TB/s** is the real boast. It is the speed at which GPUs exchange data, and it is where the 'behaves like one accelerator' property comes from. ③ **120 kW** is roughly **ten times** a conventional server rack, which typically draws 5 to 15 kW. **That power becomes heat, and air cannot remove it** — hence liquid cooling. The same figure explains why AI companies now secure **power and land before chips** (「What AI data-center power is」·「What the data-center power crunch is」). Nvidia has cited **30 times faster real-time inference for trillion-parameter models** with this system, a multiple that shifts depending on which generation it is measured against.
- Why build this at all
- **Because the bottleneck moved from the arithmetic to the wire.** There was a time when one GPU could hold one model. Then adding GPUs simply raised throughput. **A frontier model no longer fits in a single GPU's memory** (「What a frontier model is」·「What a context window is」). The model has to be split across cards, and the moment it is, **GPUs must exchange data mid-calculation.** At that point, **however fast the GPU is, a slow interconnect sets the pace of the whole system.** Compute units sit idle waiting for data. **NVL72 is not a product that made the calculator faster. It is a product that made the wire between calculators thicker.** What HBM did inside a chip — stack memory to raise bandwidth (「What HBM is」) — NVL72 does inside a rack. **And that is why this name attaches to Korea's semiconductor story.** Every Blackwell GPU carries HBM, and every rack carries 72 GPUs. **Rack demand is HBM demand.** The connection is exactly what Goldman Sachs was pricing when it held a KOSPI target of 12,000 on September 7 (「Goldman keeps its KOSPI target at 12,000 (September 7)」).
It is not a product that made GPUs faster. It is a product that made the wire between GPUs thicker. Everything about Grace Blackwell NVL72 follows from that sentence.
1. Read the name in three parts
| Part | Meaning |
|---|---|
| Grace | Nvidia's CPU |
| Blackwell | The GPU of that generation |
| NVL72 | 72 GPUs bound by NVLink |
So it names a rack, not a chip. The construction stacks:
| Layer | Composition |
|---|---|
| GB200 Superchip | 1 Grace CPU + 2 Blackwell GPUs, joined by 10 TB/s NVLink-C2C |
| Rack (NVL72) | 36 superchips → 36 CPUs, 72 GPUs |
| Fabric | Two-tier NVLink switch |
The last line is the point. Many GPUs in one cabinet is not new. What distinguishes NVL72 is that the 72 GPUs appear to software as a single accelerator — one large memory space rather than 72 cards to feed by hand.
The unit of sale moved from the card to the rack.
2. The numbers
| Item | Value |
|---|---|
| Grace CPUs | 36 |
| Blackwell GPUs | 72 |
| AI compute | about 1.4 exaflops |
| Aggregate NVLink bandwidth | about 130 TB/s |
| Power draw | about 120 kW |
| Cooling | Liquid |
They explain each other.
- 1.4 exaflops is a low-precision figure. Do not compare it with supercomputer ranking numbers.
- 130 TB/s is the real boast, and the source of the 'one accelerator' property.
- 120 kW is about ten times a conventional server rack's 5 to 15 kW. Power becomes heat, and air cannot carry it away — hence liquid cooling.
That third figure also explains why AI firms now secure power and land before chips (「What AI data-center power is」·「What the data-center power crunch is」).
3. Why this unit exists
Once, one GPU held one model, and adding GPUs raised throughput. A frontier model no longer fits in one GPU's memory (「What a frontier model is」). Split it across cards and the GPUs must exchange data mid-calculation.
| Where the bottleneck sits | Before | Now |
|---|---|---|
| What is slow | The compute units | The wire between GPUs |
| The fix | Faster chips | Thicker interconnect |
However fast the GPU, a slow wire sets the pace for everything. Compute units idle, waiting.
What HBM did inside a chip (「What HBM is」), NVL72 does inside a rack.
4. How it connects to Korea
Every Blackwell GPU carries HBM, and every rack carries 72 GPUs. Rack demand is HBM demand.
The name reached wider circulation on September 7, 2026, when Jensen Huang said GPT-6 Astra was trained on "more than 100,000 Grace Blackwell NVLink72" units (「Jensen Huang declares AGI has arrived (September 7)」), adding that 400,000 GPUs would run next.
The same day, Goldman Sachs held a KOSPI target of 12,000 on the same chain — capex buys racks, racks carry HBM (「Goldman keeps its KOSPI target at 12,000 (September 7)」·「What AI capex is」).
5. What is left and what is unverified
- Specifications are not measurements. 1.4 exaflops, 130 TB/s and 120 kW are published figures.
- The '30x inference' claim depends on the comparison baseline. Full conditions were not obtained.
- The unit in "100,000 NVLink72" is ambiguous. At rack scale that is 7.2 million GPUs; at superchip scale, roughly 1,400 racks. More than 5,000 times apart.
- Later generations are out of scope, including GB300.
- Rack pricing and shipment volumes could not be confirmed from public sources.
Sources
- NVIDIA — GB200 NVL72 product page
- Introl — GB200 NVL72 deployment: managing 72 GPUs in a liquid-cooled configuration
- FiberMall — NVIDIA GB200 NVL72 for rack-scale AI
- AI Times Korea — CoreWeave deploys Nvidia Grace Blackwell GPUs at scale
- Money Today — 'AGI has arrived,' Jensen Huang declares; GPT-6 Astra scores full marks on the suneung