DEV Community

euk ela
euk ela

Posted on

Nvidia's Gross Margin Dropped 3 Points. Here's What Vera Rubin Ramp Is Actually Burning

dev.to 2026-11-18 — Nvidia's Gross Margin Dropped 3 Points. Here's What Vera Rubin Ramp Is Actually Burning

改编说明(事实口径)

  • 源稿:掘金 2026-11-18 长文《毛利率跌 3 个点背后:Vera Rubin 爬坡到底在烧什么》(https://juejin.cn/spost/7709070757514903590),事实口径与源稿完全一致,未引入新硬事实。
  • 唯一更新:盘后涨幅口径。源稿写稿时为 CNBC 即时读数"约 3%";本轮改编采用当日 ops 日志 10:55 补核的最终读数——盘后收约 +4.8%(Finnhub 盘后报价 + Market Unclouded 20:00 ET 更新同口径),正文按此写,并在文中注明"settled around +4.8% after the initial ~3% pop"。
  • NVLink 全机架带宽口径不一(260 TB/s vs losslessnetwork 约 216 TB/s):正文沿用源稿处理,采用 72×3.6 算术一致口径并标注"Nvidia's figure"。
  • "十分之一成本"标注厂商口径;未写具体良率、未写毛利率回升时间表。

Title

Nvidia's Gross Margin Dropped 3 Points. Here's What Vera Rubin Ramp Is Actually Burning

Body

Nvidia's Q3 was a blowout: $115.8 billion in revenue, about $6.7 billion above expectations, with the data center alone doing $107.6 billion — up 110% year over year.

And yet the stock only rose about 3% in the first minutes after hours, settling around +4.8% by the after-hours close. The reason was sitting in the same earnings release: Q4 gross margin guidance of 71%–72%, roughly 3 points below this quarter's 74.9%. The company's entire explanation was one line: high initial scaling costs for the new Vera Rubin architecture.

That line deserves to be unpacked. Vera Rubin isn't a faster GPU. It's a rack-scale system — CPU, GPU, interconnect, networking, DPU, power, cooling, seven chips designed together. A generational switch here means rebuilding the whole production line at once. The ramp cost is hiding inside the words "co-designed."

What Vera Rubin actually is: the rack is the computer

Blackwell already made "the rack is the computer" real. Vera Rubin pushes it further. In its flagship NVL72 form, one rack is one machine:

Component Spec (Nvidia's figures)
Rubin GPU 72 chips, each with 288GB HBM4, up to 22 TB/s bandwidth, 50 PFLOPS NVFP4 inference
Vera CPU 36 chips, each with 88 custom Olympus cores, NVLink-C2C at 1.8 TB/s
NVLink 6 3.6 TB/s per GPU, ~260 TB/s aggregated across the rack
Full-rack compute 3.6 exaflops NVFP4 inference, 2.5 exaflops training
Total memory 20.7TB HBM4, 1.6 PB/s aggregate bandwidth; 54TB LPDDR5x system memory
Power & cooling 190–230kW, 100% liquid cooling, no air-cooled configuration
Power delivery 800V DC architecture, replacing the 48V rack power used for over a decade

Seven co-designed chips: Rubin GPU, Vera CPU, NVLink 6 switch, ConnectX-9 SuperNIC, BlueField-4 DPU, Spectrum-6 Ethernet switch — plus the Groq 3 LPU, added in March 2026, splitting inference work so Rubin handles prefill and the Groq chip handles decode.

One detail worth noting: Vera is Nvidia's first CPU with a fully custom-designed core, not a licensed Arm reference design. When even the CPU gets built from scratch, the ambition isn't single-chip performance — it's the cost structure of the whole rack. Nvidia claims an NVL72 cuts inference cost per million tokens to one-tenth of a Blackwell GB200 NVL72. That's the vendor's claim, with no independent verification yet.

Where the ramp money actually goes: four bills

Bill one: yield ramp. A single Rubin packs 336 billion transistors on TSMC 3nm, with two compute dies bonded into one GPU via NV-HBI packaging. More transistors and more complex packaging mean lower early yields — and every shippable chip carries a bigger share of wafer cost. This is the most predictable cost item in any new architecture's first year of volume production.

Bill two: HBM4. 288GB per card at 22 TB/s — nearly three times Blackwell's HBM3e at 8 TB/s. HBM4 gets its bandwidth from a wider interface, stacked 12 layers high, and it's currently the scarcest, most expensive memory on the market. A full rack holds 20.7TB of it, so memory's share of the bill of materials can only be higher than last generation's.

Bill three: liquid cooling and power retrofits. A 190–230kW rack, 100% direct liquid cooling, 800V DC power — the data center has to change with it. Cold plates, coolant loops, an 800V power chain: none of that is off-the-shelf. Nvidia even launched the DSX POD pre-integrated deployment package, claiming on-site deployment can shrink from about six months to as little as 14 days once site conditions are ready. The faster the deployment, the more the integration and debugging cost concentrates in the first few quarters.

Bill four: scale hasn't arrived yet. CoreWeave only brought the first validated full rack online on June 1, 2026, and large-scale cloud deployments are concentrated in the second half of 2026. Rubin was about 20% of Q3 data center revenue, by the company's own account — the product is still early in its ramp, fixed costs aren't spreading thin yet, and margins are under pressure by definition.

Add the four bills together and a 3-point Q4 margin dip is not surprising. The real question is whether this is a one-time ramp dip or the new normal cost of the architecture. The earnings release doesn't answer that.

Three honest limitations: what this piece can't tell you

  1. Nvidia hasn't disclosed actual yield numbers, so how big the "ramp cost" really is can only be inferred backward from margin guidance — it can't be broken down precisely.
  2. The company hasn't committed to a timeline for margin recovery. "Temporary" is the market's reading, not an official promise.
  3. The "one-tenth the cost per million tokens" claim is Nvidia's own. No independent third party has published a reproducible head-to-head test yet.

What it means if you buy compute: look at whole-rack TCO

If you rent compute or deploy inference services, Vera Rubin's real meaning isn't "a faster card" — it's that the unit of pricing has changed:

  • Per-card price is no longer the cost anchor. Whole-rack power, liquid cooling, and deployment time all belong in the TCO.
  • Inference work is split across chips (prefill vs. decode), so when you pick a setup, ask: is your workload prefill-heavy or decode-heavy?
  • 800V power and full liquid cooling mean the data centers that can even host Vera Rubin are themselves a scarce resource — regional supply will affect the price you get.

In short: the earnings numbers look great, but whether Vera Rubin turns "one-tenth the cost" from a launch-slide claim into a number on a customer's invoice will decide whether Nvidia's margin climbs back next quarter. That's the number to watch in Q4 — not the $130 billion itself, but which way the 71%–72% band moves next.

What do you think — is a new-architecture margin dip a buying opportunity, or a sign the valuation has peaked? Let's talk in the comments.

Top comments (0)