Stage 6 of 9
GPUs, servers, racks and networking systems
The unit everyone counts. Price per GPU, GPUs per rack and rack power decide what a megawatt is worth.
- Unit
- GPUs
- Sells
- accelerators assembled into servers and racks with the network that ties them into a cluster
- Buys from
- Wafers and advanced packaging · Memory and storage (HBM, DRAM, NAND)
- Lead time
- 0.75 years
- Status
- skeleton
Primer
What this stage is
The accelerator assembled into something that can be plugged in: the GPU, its server, the rack that holds many servers, and the network that ties racks into one cluster. Its unit is the GPU, the thing every plan and every press release counts.
How it works
A GPU's price is set by its maker; a server's price adds the host processors, memory, storage and assembly; a rack adds power delivery, cooling and the switches inside it; and the cluster adds the optical links and switches between racks, which for the largest systems cost as much as the GPUs they connect. Rack power decides how many racks a megawatt supports. Generations turn over every year or two, and each brings more performance per watt, which is why the useful life of a GPU is argued over.
A handful of server makers assemble most of the world's AI systems on tight schedules; their lead times, and the accelerator maker's own delivery commentary, are the shipment signal.
What to watch
- Price per GPU, per server and per rack by generation.
- GPUs per rack and rack power.
- Networking cost per GPU and the optics behind it.
- Accelerator revenue and the unit shipments it implies.
- Delivery lead times; the share taken by custom silicon and second suppliers.
How it connects
A GPU installed and powered is a GPU-hour waiting to be sold. Its price, life and power draw are the inputs to the unit economics of the next stage.
Figures
| Metric | Value | Unit | As of | Scope | Confidence | Source | Note |
|---|---|---|---|---|---|---|---|
| gb200_nvl72_rack_price_low | 2.8 | USD million per GB200 NVL72 rack (low end of reported range) | 2026-03-24 | NVIDIA GB200 NVL72 (72 GPUs) | low | Tom's Hardware (syndicated on Yahoo Finance), 'Price of Nvidia's Vera Rubin NVL72 racks skyrockets to as much as $8.8 million apiece' (24 Mar 2026) | As printed: 2.8e+06 USD per GB200 NVL72 rack (low end of reported range). Low end of a $2.8-3.4 million range attributed to an unnamed 'Tom's Hardware source with knowledge of the matter'; the same article puts an inference-optimised GB300 NVL72 at $6-6.5 million and a Vera Rubin VR200 NVL72 at $5-7 million. NVIDIA and the OEMs publish no list price, so every rack price in circulation is an estimate. Arithmetic: $2.8-3.4 million / 72 GPUs = $38,900-47,200 per GPU at the rack level. |
| hgx_b200_8gpu_server_price_low | 400,000 | USD per 8-GPU HGX B200 server (low end of reported range) | 2026-07-16 | global (OEM and marketplace quotes) | low | Mercatus AI blog, 'NVIDIA B200 Server Price in 2026: What an 8-GPU HGX System Costs' (16 Jul 2026) | Low end of a reseller's compiled $400,000-500,000 range for a complete 8-GPU HGX B200 server (CPUs, memory, NICs, storage included), based on public OEM and marketplace quotes as of July 2026; the article puts NVIDIA's own DGX B200 appliance at about $515,000 list and the bare B200 card at $30,000-40,000 list in 8-GPU volumes. Not a filing or a published price list. |
| nvl72_rack_power_total | 132 | kW per NVL72 rack (rated power-shelf capacity) | 2026-09 | Supermicro SRS-GB200-NVL72-M1 (NVIDIA GB200 NVL72) | high | Supermicro, 'Supermicro NVIDIA GB200 NVL72 Datasheet' (PDF), rack specifications table | Rated rack power from eight 33 kW power shelves for a 72-GPU GB200 NVL72 rack; the datasheet gives operating power as 125 kW to 135 kW, NVIDIA's own marketing figure is 120 kW, and Vertiv's NVIDIA-co-developed reference design is sized for 'up to 132kW per rack'. Covers GPUs, CPUs, NVLink switches and in-rack CDU but not the scale-out (InfiniBand/Ethernet) switches in separate racks. |
| dell_ai_server_quarterly_revenue | 16.4 | USD billion per quarter | 2026-07-31 | Dell Technologies | high | Dell Technologies, Form 8-K exhibit 99.1, 'Dell Technologies Delivers Second Quarter Fiscal 2027 Financial Results' (1 Sep 2026) | AI-optimised server revenue recognised in the quarter ended 31 July 2026 (fiscal Q2 2027), roughly double a year earlier; Dell raised its fiscal-2027 AI server revenue outlook to $74.0 billion in the same release. |
| dell_ai_server_backlog | 95 | USD billion of unshipped AI server orders | 2026-07-31 | Dell Technologies | high | Dell Technologies, Form 8-K exhibit 99.1, 'Dell Technologies Delivers Second Quarter Fiscal 2027 Financial Results' (1 Sep 2026) | AI server backlog at quarter end after booking $60.9 billion of AI server orders in the quarter; arithmetic: $95 / $16.4 = 5.8 quarters of shipments at the current rate, a direct measure of orders outrunning what can be built and shipped. |
| supermicro_fy2026_net_sales | 39.1 | USD billion per fiscal year | 2026-06-30 | Super Micro Computer | high | Super Micro Computer, press release 'Supermicro Announces Fourth Quarter and Full Fiscal Year 2026 Financial Results' (11 Aug 2026) | Full fiscal-year net sales (year ended 30 June 2026), up 78% from $22.0 billion; Supermicro does not split out AI-server revenue but guides fiscal 2027 to $65.0-72.0 billion and says it booked more than $60 billion of new orders in the year. |
| broadcom_ai_semiconductor_quarterly_revenue | 16.7 | USD billion per quarter | 2026-08-02 | Broadcom | high | Broadcom, Form 8-K exhibit 99.1, 'Broadcom Inc. Announces Third Quarter Fiscal Year 2026 Financial Results and Quarterly Dividend' (2 Sep 2026) | AI semiconductor revenue in the quarter ended 2 Aug 2026 (fiscal Q3 2026); this combines custom AI accelerators (XPUs) and AI networking silicon (Tomahawk/Jericho Ethernet switches, optics), which Broadcom does not split in the release; guided to $21.7 billion for fiscal Q4. |
| ethernet_share_of_ai_backend_switch_sales | 66.7 | % of data-centre switch sales in AI clusters (stated as 'about two-thirds') | 2026-03 | global | medium | Dell'Oro Group press release, 'Ethernet Extends Lead in AI Scale-Out Networks Despite Strong InfiniBand Rebound' (2 Jun 2026), covering Q1 2026 | Dell'Oro states the share only as 'about two-thirds' (66.7 is our decimal of that phrase, not a published figure); the remaining roughly one-third is InfiniBand, whose sales 'more than tripled' in the quarter on NVIDIA Blackwell Ultra shipments, so both fabrics are growing fast while Ethernet gains share; 800 Gbps switches were 'the vast majority' of Ethernet shipments and 1.6 Tbps is sampling for a 2H26 ramp. |
| vertiv_backlog | 15 | USD billion of unshipped orders | 2025-12-31 | Vertiv | high | Vertiv Holdings, Form 8-K exhibit 99.1, 'Vertiv Reports Strong Fourth Quarter' (11 Feb 2026) | Backlog for the largest listed vendor of data-centre power and thermal (including liquid-cooling) equipment at year-end 2025, more than a full year of its 2026 sales guidance ($14,000 million); the Q2 2026 release (29 Jul 2026) did not restate backlog, and Q4 2025 book-to-bill was '~2.9x'. |
| servers_share_of_ai_datacenter_tco | 60 | % of annualised total cost of ownership of a 1 GW AI data centre | 2026-05-14 | global (model: 1 GW of GB200 NVL72) | medium | Epoch AI, data insight 'Total cost of ownership of a one-gigawatt AI data center' (14 May 2026) | Epoch's bottom-up model puts servers (GPU racks) at $5 billion of an $8.5 billion annualised TCO for a 1 GW facility built entirely of GB200 NVL72; network infrastructure is a separate line of about $1.2 billion a year, i.e. roughly 14% of TCO and about a quarter of the server line, which is the best public proxy found for networking's share of AI-cluster cost. |
| liquid_cooling_penetration_ai_chips | 53 | % of AI chips shipped in 2026 that are liquid-cooled (TrendForce estimate) | 2026-08-17 | global | medium | TrendForce press release, 'Liquid Cooling Becomes Standard for High-End AI Infrastructure, with Penetration Expected to Reach 53% in 2026' (17 Aug 2026) | TrendForce's estimate for calendar 2026 (published mid-year, so partly actual and partly forecast), rising toward 60% in 2027; the release attributes the shift to chip TDPs above 1 kW and rack-scale systems drawing 'hundreds of kilowatts', and says NVIDIA rack-scale shipments are expected to double in 2026. Uptime Institute's operator survey (registration-walled) reportedly shows a much lower share of sites, so this is a share of new AI chips, not of data centres. |
| ai_server_shipment_growth_2026 | 31 | % year-on-year growth in AI server unit shipments, 2026 (TrendForce estimate, 'nearly 31%') | 2026-08-03 | global | medium | TrendForce press release, 'AI Server Shipments Forecast Raised to Nearly 31% YoY in 2026 as 90% Surge in CSP CapEx Fuels Infrastructure Expansion' (3 Aug 2026) | Unit-shipment growth for AI servers (GPU and ASIC systems counted as servers, not racks), revised up from 28% during the year as the nine largest cloud providers' combined capex was projected up about 90% in 2026; dollar growth is far higher because rack-scale NVL72 systems carry many more GPUs per server than 8-GPU boxes. |
Conversions
| From → To | Factor | Unit | As of | Confidence | Source | Note |
|---|---|---|---|---|---|---|
| Data centre shell and cooling → GPUs, servers, racks and networking | 120 | kW of IT load per liquid-cooled GB200 NVL72 rack | 2024-10 | high | NVIDIA Technical Blog, NVIDIA Contributes NVIDIA GB200 NVL72 Designs to Open Compute Project (15 Oct 2024) | Design figure; 1 MW of IT load supports about 8 such racks (1,000 / 120 = 8.3); Microsoft reports ~140 kW per rack in production for GB200/GB300, which gives about 7 racks per MW. |
| Data centre shell and cooling → GPUs, servers, racks and networking | 72 | GPUs per GB200 NVL72 rack | 2024-10 | high | NVIDIA, GB200 NVL72 product page | 72 Blackwell GPUs plus 36 Grace CPUs per rack; at 120 kW per rack that is 120 / 72 = 1.67 kW per GPU at the rack (including CPUs, NVLink switches and fans but excluding facility cooling and power losses, which the PUE factor adds). |
| Data centre shell and cooling → GPUs, servers, racks and networking | 479 | Blackwell GPUs per MW of campus IT load, all-in (compute, storage, networking) | 2026-09-24 | medium | Epoch AI, AI Data Centers directory (estimates; page updated 24 Sep 2026) | Arithmetic from Epoch's Abilene estimate: 100.8k B200 + 100.8k B300 = 201,600 GPUs / 421 MW IT = 479 GPUs per MW, i.e. 2.09 kW of IT load per GPU; Epoch's Colossus 2 figures give a similar 440,000 / 946 = 465 per MW; SemiAnalysis counted ~110k GB200 for ~200 MW at Colossus 2 (550 per MW); these are tracker estimates, not operator disclosures. |
| GPUs, servers, racks and networking → Memory and storage (HBM, DRAM, NAND) | 80 | GB of HBM per GPU (NVIDIA H100 SXM, HBM3) | 2026-09-27 | high | NVIDIA, H100 Tensor Core GPU product page specification table | H100 SXM lists 80GB of HBM3 at 3.35TB/s; the PCIe-based H100 NVL variant lists 94GB. |
| GPUs, servers, racks and networking → Memory and storage (HBM, DRAM, NAND) | 5 | HBM stacks per GPU (NVIDIA H100 SXM5, active HBM3 stacks) | 2022-03-22 | high | NVIDIA Technical Blog, 'NVIDIA Hopper Architecture In-Depth', March 22, 2022 | Five active 16 GB HBM3 stacks give 80 GB; the package has six stack sites, one left unused for yield. |
| GPUs, servers, racks and networking → Memory and storage (HBM, DRAM, NAND) | 141 | GB of HBM per GPU (NVIDIA H200 SXM and H200 NVL, HBM3e) | 2026-09-27 | high | NVIDIA, H200 Tensor Core GPU product page and specification table | Per-GPU HBM3e capacity from NVIDIA's own specification table (4.8TB/s); same GPU die as H100 with six 24 GB stacks, of which 141 GB is exposed. |
| GPUs, servers, racks and networking → Memory and storage (HBM, DRAM, NAND) | 180 | GB of HBM3e per GPU as configured in DGX B200 (1,440 GB / 8 GPUs) | 2026-09-27 | high | NVIDIA, DGX B200 product page specification table | Arithmetic: 1,440 GB total across 8x Blackwell GPUs = 180 GB per GPU as shipped; the Blackwell GPU's full on-package capacity is 192 GB (next row). |
| GPUs, servers, racks and networking → Memory and storage (HBM, DRAM, NAND) | 192 | GB of HBM3e per GPU (NVIDIA Blackwell B200, on-package capacity) | 2025-08-22 | medium | NVIDIA Technical Blog, 'Inside NVIDIA Blackwell Ultra', Aug 22, 2025 | Derived: Blackwell Ultra's 288 GB is stated to be 50% more than Blackwell, so 288 / 1.5 = 192 GB for Blackwell, i.e. eight 8-Hi stacks of 24 GB. |
| GPUs, servers, racks and networking → Memory and storage (HBM, DRAM, NAND) | 288 | GB of HBM3e per GPU (NVIDIA Blackwell Ultra / GB300) | 2025-08-22 | high | NVIDIA Technical Blog, 'Inside NVIDIA Blackwell Ultra', Aug 22, 2025 | Per-GPU HBM3e capacity for the GB300 generation; a GB300 NVL72 rack lists 72 GPUs and 20 TB of GPU memory (72 x 288 GB = 20,736 GB). |
| GPUs, servers, racks and networking → Memory and storage (HBM, DRAM, NAND) | 8 | HBM stacks per GPU (NVIDIA Blackwell Ultra, 12-Hi HBM3e stacks of 36 GB) | 2025-08-22 | high | NVIDIA Technical Blog, 'Inside NVIDIA Blackwell Ultra', Aug 22, 2025 | Eight stacks x 12 dies x 3 GB = 288 GB; stack count per GPU has gone 5 (H100) to 6 (H200) to 8 (B200 and GB300), and dies per stack from 8 to 12. |
| GPUs, servers, racks and networking → Memory and storage (HBM, DRAM, NAND) | 72 | GPUs per GB300 NVL72 rack, with 20 TB of GPU memory per rack | 2026-09-27 | high | NVIDIA, GB300 NVL72 product page specification table | The page lists 72 Blackwell Ultra GPUs and 20 TB of GPU memory per rack; 72 x 288 GB = 20.7 TB, so '20 TB' is a rounded rack total. |
| Wafers and advanced packaging → GPUs, servers, racks and networking | 3,958 | teraFLOPS FP8 Tensor Core per GPU, with sparsity (H100 SXM) | 2026-09 | high | NVIDIA, H100 Tensor Core GPU product page, specifications table | Hopper-generation peak FP8 throughput per GPU as listed in the datasheet table; the figure is 'with sparsity', so dense FP8 is half (1,979 teraFLOPS), and FP16/BF16 with sparsity is 1,979 teraFLOPS. The quote is a table cell, hence shorter than eight words. |
| Wafers and advanced packaging → GPUs, servers, racks and networking | 9 | petaFLOPS FP8 Tensor Core per GPU, with sparsity (B200, from 72 PFLOPS per 8-GPU DGX B200) | 2026-09 | high | NVIDIA, DGX B200 product page, specifications table | Arithmetic: DGX B200 lists FP8 Tensor Core 72 PFLOPS for 8 Blackwell GPUs, so 72 / 8 = 9 PFLOPS per B200, about 2.3x the H100's 3.958 PFLOPS; the GB200 NVL72 page lists FP4 1,440 PFLOPS for 72 GPUs, i.e. 20 PFLOPS FP4 per GB200 GPU. The quote is a table cell. |
| Wafers and advanced packaging → GPUs, servers, racks and networking | 72 | GPUs per rack (GB200 NVL72) | 2026-09 | high | NVIDIA, GB200 NVL72 product page, specifications table | One NVL72 rack holds 36 Grace CPUs and 72 Blackwell GPUs in a single NVLink domain; the same table lists 13.4 TB of HBM3E per rack, i.e. 13,400 / 72 = 186 GB of HBM per GPU (or 372 GB per two-GPU superchip), which is the bridge to the memory stage. |
| Wafers and advanced packaging → GPUs, servers, racks and networking | 25 | gross CoWoS interposers (GPU packages) per 300 mm wafer, before yield (low end of 25-28) | 2026-04-13 | low | Silicon Analysts, CoWoS Cost 2026: $750-1,100/Chip, Chiplet vs Monolithic (Apr. 13, 2026) | Analyst-site estimate with no primary citation, for a B200-class package (two near-reticle dies plus eight HBM stacks); applying its 60-70% interposer yield gives about 15-19 good packages per CoWoS wafer, so 130,000 wafers per month would be roughly 2-2.5 million such GPUs per month. TSMC does not publish packages per wafer; treat as an order-of-magnitude figure. |
| GPUs, servers, racks and networking → Data centre shell and cooling | 72 | GPUs per NVL72 rack | 2026-09 | high | NVIDIA, GB200 NVL72 product page (specifications table) | One GB200 or GB300 NVL72 rack holds 72 Blackwell GPUs and 36 Grace CPUs across 18 compute trays (4 GPUs + 2 CPUs each) plus 9 NVLink switch trays; a Vera Rubin NVL72 rack keeps the same 72-GPU count. |
| GPUs, servers, racks and networking → Data centre shell and cooling | 8 | GPUs per HGX/DGX 8-GPU server | 2026-09 | high | NVIDIA, DGX B200 product page (specifications table) | The HGX baseboard design used by every OEM 8-GPU server (Supermicro, Dell XE9680, HPE, etc.) carries eight GPUs; a DGX B200 occupies 10 rack units, so air-cooled racks typically hold 2 to 4 such servers (16 to 32 GPUs) against 72 in a liquid-cooled NVL72. |
| GPUs, servers, racks and networking → Data centre shell and cooling | 1.83 | kW per GPU at the rack (NVL72, rated rack power incl. CPUs, NVLink switches, in-rack CDU) | 2026-09 | high | Supermicro, 'Supermicro NVIDIA GB200 NVL72 Datasheet' (PDF), rack specifications table | Arithmetic: 132 kW rated rack power / 72 GPUs = 1.833 kW per GPU; using the datasheet's operating range of 125-135 kW gives 1.74-1.88 kW per GPU. Excludes scale-out network switches, storage and facility cooling overhead (PUE), which belong to the datacenter stage. |
| GPUs, servers, racks and networking → Data centre shell and cooling | 1.79 | kW per GPU at the server (HGX/DGX B200 8-GPU, max system power incl. CPUs and NICs) | 2026-09 | high | NVIDIA, DGX B200 product page (specifications table) | Arithmetic: 14.3 kW max system power / 8 GPUs = 1.79 kW per GPU, including the two CPUs, memory, storage and ConnectX-7/BlueField-3 NICs inside the server but not the external switches; per GPU this is close to the NVL72 figure despite the NVL72's far higher density per rack. |
| GPUs, servers, racks and networking → Memory and storage (HBM, DRAM, NAND) | 186 | GB of HBM3E per GPU (GB200) | 2026-09 | high | NVIDIA, GB200 NVL72 product page (specifications table) | Arithmetic: 13.4 TB of HBM3E per rack / 72 GPUs = 0.186 TB = 186 GB per GPU (from NVIDIA's rounded rack total; each Blackwell GPU carries 8 HBM3E stacks, so about 23 GB per stack). |
| GPUs, servers, racks and networking → Memory and storage (HBM, DRAM, NAND) | 278 | GB of HBM3e per GPU (GB300 / Blackwell Ultra) | 2026-09 | high | NVIDIA, GB300 NVL72 product page (specifications table) | Arithmetic: 20 TB of GPU memory per GB300 NVL72 rack / 72 GPUs = 0.278 TB = 278 GB per GPU, a 49% increase in HBM per GPU over GB200 for the same 72-GPU rack, which is why HBM demand grows faster than GPU unit counts. |
| GPUs, servers, racks and networking → GPU-hours | 56,000 | USD per GPU deployed in an 8-GPU HGX B200 server (system-level price, all components) | 2026-07-16 | low | Mercatus AI blog, 'NVIDIA B200 Server Price in 2026: What an 8-GPU HGX System Costs' (16 Jul 2026) | Arithmetic check: a typical $450,000 server / 8 GPUs = $56,250 per GPU, versus a bare B200 card list price of $30,000-40,000, so the server wrapper (CPUs, memory, NICs, storage, chassis, integrator margin) adds roughly 40-90% on top of the GPU; the $2.8-3.4 million GB200 NVL72 estimate implies $38,900-47,200 per GPU at rack level. Reseller compilation, not a filing. |
Players
| Company | Ticker | Role | Listed | Note |
|---|---|---|---|---|
| Hon Hai Precision Industry (Foxconn) | TWSE:2317 | largest AI server and NVL72 rack assembler (ODM/EMS) | yes | Q2 2026 revenue reached NT$2.53 trillion; the company said full-year AI rack shipments 'are also expected to more than double', with Vera Rubin racks entering mass production in Q3 2026, and secondary coverage put cloud and networking at 51% of revenue, passing half for the first time. |
| Quanta Computer | TWSE:2382 | AI server ODM for hyperscalers (GB200/GB300 racks, custom-ASIC servers) | yes | Servers were 80% of Q1 2026 revenue of NT$809.22 billion and AI servers 75% of server revenue, but gross margin fell to 4.78%, showing how thin ODM margins are on NVIDIA rack systems. |
| Wiwynn | TWSE:6669 | hyperscale server and rack ODM (Wistron group) | yes | Q1 2026 consolidated revenue of NT$276.508 billion, up 62.0% year on year, with profit after tax of NT$14.114 billion; one of the four ODMs building GB200/GB300 NVL72 racks. |
| Super Micro Computer | NASDAQ:SMCI | AI server and liquid-cooled rack OEM (Data Center Building Block Solutions) | yes | Fiscal 2026 net sales of $39.1 billion (up from $22.0 billion), Q4 net sales $11.1 billion, more than $60 billion of new orders in the year and fiscal 2027 guidance of $65.0-72.0 billion. |
| Dell Technologies | NYSE:DELL | AI server OEM (PowerEdge XE, NVL72 racks) for enterprises and neoclouds | yes | Booked $60.9 billion of AI server orders and recognised $16.4 billion of AI server revenue in the quarter ended 31 Jul 2026, with a $95 billion AI backlog; traditional servers and networking revenue was $10.5 billion, up 122%. |
| Hewlett Packard Enterprise | NYSE:HPE | AI systems and server OEM (Cray, ProLiant, GB200 NVL72 by HPE), plus Juniper networking | yes | Fiscal Q3 2026 (quarter to 31 Jul 2026) revenue of $12.2 billion, up 34%, with Server revenue of $6.8 billion, up 35.3%, and order backlog 'at a record level'; HPE's own spec sheet rates the GB200 NVL72 rack at 132 kW. |
| Arista Networks | NYSE:ANET | Ethernet switching for AI back-end and front-end networks | yes | Q2 2026 revenue of $3.036 billion, up 37.7% year on year and its first $3 billion quarter, with Q3 guided to about $3.3 billion and new 1.6 Tbps liquid-cooled AI fabric platforms; does not disclose AI revenue separately. |
| Cisco Systems | NASDAQ:CSCO | Ethernet switching, optics and Silicon One for hyperscaler AI back-end networks | yes | Took $4 billion of hyperscaler AI-infrastructure orders in fiscal Q4 2026, bringing fiscal 2026 to $9.3 billion, delivered about $4 billion of AI-infrastructure revenue in FY2026 and expects $7.5 billion in FY2027; networking product orders grew 40% in Q4. |
| Broadcom | NASDAQ:AVGO | Ethernet switch silicon (Tomahawk/Jericho), optics and custom AI accelerators | yes | AI semiconductor revenue of $16.7 billion in fiscal Q3 2026 (up 221%), guided to $21.7 billion in Q4; the release says demand for custom AI accelerators and networking 'continues to be very strong'. |
| Coherent | NYSE:COHR | optical transceivers (800G/1.6T) and lasers for AI data centres | yes | Fiscal Q4 2026 (quarter to 30 Jun 2026) revenue of $2.05 billion, up 34% year on year, with Datacenter & Communications revenue of $1,615.0 million and Q1 FY2027 guided to $2.2-2.4 billion; peers Lumentum ($1.01 billion, up 109.3%) and Credo ($479.0 million, up 114.7%) grew faster from smaller bases. |
| Vertiv | NYSE:VRT | power distribution, thermal and liquid-cooling (CDUs) for AI racks | yes | Backlog of $15.0 billion at end-2025 (up 109%) against 2026 sales guidance of $14,000 million; Q2 2026 net sales $3,274 million, up 24%; co-developed with NVIDIA a 7 MW GB200 NVL72 reference design at 'up to 132kW per rack'. Main rival Schneider Electric (Motivair) sells CDUs from 105 kW to 2.5 MW. |
| CoolIT Systems (being acquired by Ecolab) | private (Ecolab NYSE:ECL after closing) | direct-to-chip cold plates and CDUs for GPU servers and NVL72 racks | no | Ecolab agreed on 20 Mar 2026 to pay approximately $4.75 billion in cash for CoolIT, which is expected to generate approximately $550 million in sales over the following 12 months, implying about 8.6x sales for a cold-plate and CDU maker; closing expected in Q3 2026. |
Signals
| Date | Kind | Actor | Counterparty | Claim | Value | Confidence | Source |
|---|---|---|---|---|---|---|---|
| 2026-09-01 | contract | Dell Technologies | AI server orders of $60.9 billion in the quarter to July 31, 2026 against $16.4 billion shipped; AI server backlog of $95 billion, about 5.8 quarters of shipments at the current rate | 95 USD bn | confirmed | Dell Technologies Q2 FY2027 earnings release (8-K exhibit 99.1, 2026-09-01) |