Nvidia Hopper Explained: Why a 2022 GPU Is Still the Smart Money in 2026

Nvidia Hopper is the GPU architecture behind the H100, H200, and GH200. Announced in March 2022, it remains the largest deployed AI accelerator base in history, with cumulative shipments measured in millions of units. In 2026, Hopper occupies the value tier of AI infrastructure: H100 rentals have fallen 64–75% from their 2023 peak to roughly $2.29–3.12/hour, used H100s trade at $6,000–15,000, and the entire platform runs in air-coolable 20–45 kW racks that ordinary facilities can actually host.

This post covers the Hopper lineup → what it demands from a facility → why it still matters in 2026 → the Hopper-vs-Blackwell decision → what a 1 MW Hopper deployment looks like → Hopper's role in European sovereign AI.

The Hopper lineup, in buyer's terms

Hopper, named after Grace Hopper, launched at GTC in March 2022 and hit full production that September. Five products matter to an infrastructure buyer.

H100 SXM is the flagship: 80 GB HBM3 at 3.35 TB/s, just under 4 PFLOPS of FP8 with sparsity, up to 700 W, with 900 GB/s NVLink connecting GPUs on an 8-way HGX baseboard. H100 PCIe is the 300–350 W add-in card for mainstream servers: less bandwidth, far easier hosting. H100 NVL is the paired-card inference variant with 94 GB per GPU. H200, shipping from Q2 2024, keeps the same compute die but jumps memory to 141 GB of HBM3e at 4.8 TB/s: 1.76x the capacity, 1.4x the bandwidth. For serving large models, memory is the binding constraint, which makes H200 the most underrated part in the family. And GH200 fuses a 72-core Grace CPU to a Hopper GPU at 900 GB/s; it powers Europe's first exascale machine.

There's also the H20, the export-compliant China variant that triggered a $5.5 billion Nvidia write-down during the 2025 export-control whiplash. Three years after launch, Hopper derivatives were still strategically significant silicon. That tells you something about the architecture's staying power.

What Hopper asks of a facility (refreshingly little)

Here's the infrastructure profile, straight from Nvidia's own documentation.

An 8-GPU DGX H100 server draws about 10.2 kW maximum. Nvidia's DGX SuperPOD design guide (2023) caps racks at four systems (roughly 41 kW) and recommends two per rack (~20–25 kW) in legacy air-cooled rooms. So real-world Hopper lives in the 20–45 kW band.

That band is the entire point. It's air-coolable. Hot-aisle containment handles it; rear-door heat exchangers give you headroom; Schneider Electric published guidance as recently as January 2026 on running H100/H200-class GPUs on air. No facility water loop. No CDU plant. No 1.4-tonne rack on a floor that can't take it. Compare that with what Blackwell's GB200 NVL72 demands, meaning 120–132 kW liquid-cooled racks, and you see why Hopper marks the boundary between "facility upgrade" and "new building."

Networking follows Nvidia's reference architecture: InfiniBand NDR 400 Gb/s, eight ConnectX-7 adapters per server, separate compute, storage, and management fabrics. Mature, documented, and your integrator has built it before. Don't underrate boring.

Why Hopper still matters in 2026

Most coverage treats Hopper as yesterday's chip. The market disagrees, loudly.

The installed base is enormous. Omdia estimated ~300,000 H100s shipped in Q2 2023 alone (the famous "900 tons" figure), with Meta and Microsoft each absorbing roughly 150,000 that year. Millions of Hopper GPUs are deployed and working.

Prices collapsed, which created the opportunity. H100 rental fell from $8+/hour at the 2023 peak (some clouds charged over $12) to a 2026 median around $2.29–3.12/hour, with budget providers under $2. On the buy side, used H100 SXM units trade at $6,000–15,000, with complete used 8-GPU servers at $150,000–180,000 (Hashrate Index, 2026). The chip didn't get worse. It got cheap.

And demand caught the falling knife. Rental prices firmed roughly 18% off their mid-2025 lows as inference demand absorbed the surplus (Silicon Data, 2026). CoreWeave reported that H100 capacity coming off expiring 2022-era contracts re-booked immediately at ~95% of original pricing; its 2020-vintage A100 fleet is still fully booked. That's the empirical answer to the depreciation doomers: older GPUs don't die, they cascade into inference. The accounting fight over whether GPUs live 3 years or 6 is real (Michael Burry says 2–3; Amazon shortened server lives while Meta extended them, in the same quarter). The fleet data so far sides with the optimists.

Who buys Hopper in 2026? Anyone running inference or fine-tuning on models in the 7B–70B class that dominates enterprise AI, anyone capacity-constrained on capital, and anyone whose facility can deliver 40 kW racks but not 130 kW ones. Which is most facilities. The economics of the rental side (who rents these GPUs out, and how the math works) are covered in how the neocloud business works.

Hopper vs Blackwell: the actual decision

At frontier scale, Blackwell wins and it isn't close. Nvidia's MLPerf-based analysis puts GB200 NVL72 at roughly 2x training performance per dollar versus H100, and the gains on giant-model inference are larger still.

But "performance per dollar" assumes you're paying list price for both and can host both. Flip to performance per capital dollar with real 2026 prices, and the calculus inverts for most workloads. A used H100 server at $160k against a $3.9M all-in NVL72 rack is a 24x capital gap. If your models fit in 80–141 GB per GPU and your traffic doesn't need a 72-GPU coherence domain, Hopper delivers the tokens at a fraction of the capital and runs in a facility you can stand up this year.

Sources: Nvidia (2023–25); Hashrate Index (2026); Silicon Data (2026); SemiAnalysis (2025).

The pragmatic industry view, voiced by CoreWeave's CEO at GTC 2026: leading-edge silicon takes frontier training, older GPUs cascade down to inference. The fleet is a waterfall, not a scrapheap. And with Rubin raising rack power again in H2 2026, the waterfall only gets taller.

What a 1 MW Hopper deployment actually looks like

Worked example, from Nvidia's own figures. At ~10.2 kW per HGX H100 server, a 1 MW IT load supports roughly 80 servers (about 640 GPUs), or closer to 70 servers (~560 GPUs) once you reserve 10–15% for fabric, storage, and management. At Nvidia's recommended four servers per rack, that's 18–20 compute racks at ~41 kW each, plus networking and storage rows.

Every kilowatt of that envelope is achievable with contained air cooling or hybrid rear-door exchangers. Which is precisely why Hopper-class density is the sweet spot for enterprise, edge, and factory-built modular deployments: 20–45 kW racks fit prefabricated power and cooling modules that ship today, with the cooling architecture chosen to match the climate and load rather than dictated by the silicon. The full sizing logic, translating AI servers into rack power, cooling, and module design, is in our edge AI infrastructure guide, and the end-to-end build picture in how to build an AI data center.

The European angle: Hopper is sovereign AI's workhorse

Europe's flagship compute runs on Hopper. JUPITER at Forschungszentrum Jülich, Europe's first exascale supercomputer, inaugurated September 2025, is built on roughly 24,000 GH200 Grace Hopper superchips. The EuroHPC AI Factories program is rolling out 19 AI factories across the continent, several anchored on Hopper-class silicon, and the October 2025 round selected a site in Czechia.

For national and regional AI programs without the budget or grid position for liquid-cooled rack-scale systems, Hopper-class density is what's deployable now: on air, in months, in facilities that can be manufactured rather than constructed.

The uncomfortable truth about AI infrastructure in 2026 is that the most rational GPU for most organizations is the one the headlines stopped covering two years ago. Hopper is cheap, proven, air-coolable, and in stock. Sometimes the smart money buys the boring thing.

FAQ

What is Nvidia Hopper?

Nvidia Hopper is the GPU architecture announced in March 2022, named after computing pioneer Grace Hopper. It includes the H100 (80 GB HBM3), H200 (141 GB HBM3e), H100 NVL inference variant, and GH200 Grace Hopper superchip. It preceded Blackwell and remains the largest deployed AI accelerator base in history.

What is the difference between H100 and H200?

Both use the same Hopper compute die with identical FP8/FP16 throughput. The H200 raises memory from 80 GB HBM3 at 3.35 TB/s to 141 GB HBM3e at 4.8 TB/s (1.76x the capacity and 1.4x the bandwidth), which materially improves large-model inference, where memory is the binding constraint.

How much does an Nvidia H100 cost in 2026?

New units list around $25,000–40,000 depending on variant, but used H100 SXM modules trade at $6,000–15,000, and complete used 8-GPU servers at roughly $150,000–180,000 (Hashrate Index, 2026). On-demand rental runs about $2.29–3.12 per GPU-hour at the market median.

Does Nvidia Hopper need liquid cooling?

No. DGX/HGX H100 systems are air-cooled as standard. Nvidia's SuperPOD design guide specifies up to four 10.2 kW servers per rack (~41 kW) with air cooling and containment; rear-door heat exchangers add headroom. This is the key facility difference from Blackwell's GB200 NVL72, which requires direct-to-chip liquid cooling.

Is Hopper still worth buying in 2026?

For inference and fine-tuning of models in the 7B–70B class, usually yes. Hopper delivers the best performance per capital dollar at 2026 prices, runs in air-cooled 20–45 kW racks that existing and modular facilities can host, and CoreWeave reports expiring H100 capacity re-booking at ~95% of original rates.

How many GPUs fit in a 1 MW Hopper deployment?

Roughly 560–640 H100 GPUs: about 70–80 HGX servers at ~10.2 kW each, arranged 4 per rack across 18–20 compute racks at ~41 kW, with 10–15% of the load reserved for networking, storage, and management. All of it air-coolable.

What replaced Nvidia Hopper?

Blackwell (B200/GB200, 2024–25) and Blackwell Ultra (GB300, 2025), with Vera Rubin entering volume production in H2 2026. Each generation raises rack power: ~40 kW for Hopper, 120–132 kW for GB200 NVL72, and roughly 190–230 kW for Rubin-class racks.

‍