Budgets and designs run ahead of power, memory and tools
Read together, the fortnight's chip launches, results and deals point one way. Five conclusions, each with its status as of Oct 23, 2026.
Jensen Huang said Vera Rubin is "in full production" on Oct 21 vendor statement. Google Cloud A5X is announced with GA due in November, against a target of up to 960,000 Rubin GPUs. DSX POD cuts site deployment to as little as 14 days but ships only from early 2027. Every Rubin performance figure, including MLPerf results of up to 3.74x over GB300 NVL72, is a preview submission or a vendor claim.
Microsoft announced Maia 300 (288 GB HBM4, 18 TB/s) and committed to multi-gigawatt Maia deployments from 2027. OpenAI's Jalapeño moves to small-volume deployment in late 2026. AMD's Helios is shipping and Helion partner systems are due in November; Intel's 350 W GPU D270 is sampling. Broadcom targets about $115B of AI revenue in fiscal 2027. None of the efficiency claims has been measured independently.
Meta narrowed its 2026 capex range upward to $135–145B on Oct 21, putting four-company 2026 guidance at $740–760B against $413B in 2025. Meta's Q3 free cash flow turned negative. CoreWeave fell 15.46% on Oct 14 after a Gartner neocloud warning; Microsoft dropped Nscale's Loughton site over grid delays, while Anthropic and Nscale were reported to be negotiating a second ~$45B deal.
By Epoch AI's Oct 15 estimate no tracked AI campus has yet crossed 1 GW of operating IT power; the largest, Colossus 2, draws about 946 MW. Of the 92 GW of data-center load seeking ERCOT access, 54 GW is still only proposed. Builders are reserving gas turbines through 2034, as in Venture Global's framework for up to 16 × 720 MW power islands, while Texas has paused data-center permits.
The H100 on-demand median reached $3.81 per GPU-hour, up 13% in four weeks; the B200 list median stood at $8.19. ASML said its planned 2027–28 supply "will not fully meet the demand of our customers". Samsung calls its 2026 HBM4 volume effectively sold out, CoWoS is reported sold out through 2026, and TSMC's HPC platform reached 68% of a $46.7B quarter.
Vera Rubin is in production but reservation-only until November
What Nvidia announced and measured between Oct 9 and Oct 22, and what can actually be bought today.
From full production to first tokens
Huang told GTC Berlin on Oct 21 that Vera Rubin is "in full production". The first rentable capacity is still ahead: cloud GA in November, factory-built pods from early 2027.
Google Cloud A5X Announced
GA "next month" (November 2026). Plan of up to 960,000 Rubin GPUs, about 13,333 NVL72 racks: a plan, not an installed base.
DSX POD Early 2027
Reference design built by 17 partners. Deployment at a prepared site falls from six months to as little as 14 days.
Realistic planning number
In an interview Huang put equipment arrival to token production at about one month, down from four to five.
Dated timeline
Sep 16 (background)
MLPerf Inference v6.1: Vera Rubin NVL72 in the preview category
Preview resultOct 9
vLLM support in nightly containers (CUDA 13.4, PyTorch 2.15)
Preview (nightly)Oct 21
GTC Berlin keynote: "in full production"; DSX POD; A5X announced
Vendor statement AnnouncedOct 21
DSX OS: Config Manager, Rack Management Service, Cluster Readiness Engine
Developer previewVera Rubin NVL72 vs GB200 NVL72
Rack: 36 Vera CPUs with 72 Rubin GPUs; 100% liquid cooling at coolant temperatures up to 45°C.
Headline claim: "up to 30x more work per watt" for AI agents vendor claim Not measured independently.
Ratios: vLLM blog co-authored with NVIDIA, citing NVIDIA spec pages.
MLPerf preview: Vera Rubin NVL72 vs GB300 NVL72
Preview, not GA preview
Neither hardware nor software is the generally available configuration. The round drew 30 organisations, 120 systems and 486 results.
The 3.7x headline
It is the best of six rows; the offline rows average about 1.8x. Nvidia reported 99% scaling efficiency for 288 GPUs across four racks.
vLLM day-0 (Oct 9)
AgentX with MiniMax M3: 7.84x GB200 per-GPU throughput at matched interactivity, 5.18x under a 150 tokens/s cap. Collaborators' own figures, not reproduced.
Source: Nvidia's figures as reported by shattered.io; Nebius separately posted 591,368 DeepSeek-R1 tokens/s in server mode.
DSX POD: what the 14 days include
Factory-built, not a product Nvidia sells
Partners (including Bull, Dell, HPE, Lenovo, Schneider Electric and Vertiv) finish cabling, power, liquid-cooling piping and validation before shipment. No per-POD rack count, capacity or price published.
Caveat: the facility is not included
The 14 days exclude building the facility, usually the longer task. Treat 14 days as the best case for site connection and about a month as the planning number.
Efficiency results so far
Nvidia: up to 30% less power for the same output. Lambda PoC on HGX B200: +24% throughput. Nscale, four GB300 NVL72 racks in 264 kW: +49% inference throughput.
Who is buying
Google Cloud, the anchor buyer
A5X combines NVL72, ConnectX-9 SuperNICs, the Virgo Network and fourth-generation liquid cooling. First customer: Ineffable Intelligence; WPP ran a former multi-node workload on one instance.
Weaker or secondary signals
Nscale's 66,000+ Rubin GPUs for Microsoft in Portugal (Aug 7). Spheron lists Rubin only as reservation. Dell AI Factory customers 6,500+, up from 5,000. ByteDance, Alibaba and Tencent reportedly lining up $35B of Nvidia orders, not Rubin-specific.
Our judgement
Treat Vera Rubin as reservation-only through end-2026. Size near-term plans on Blackwell and check the 5–8x vendor ratios on your own workload.
Inference silicon splits into four camps, mostly on vendor claims
AMD and Intel used the OCP Global Summit (15,000 attendees, San Jose, Oct 12–15) to attack the gaps around Vera Rubin, while Microsoft and OpenAI put dates and volumes on their own inference chips.
Helios Shipping
72-GPU OCP rack built with Meta: MI450 GPUs, EPYC Venice CPUs, Pensando networking, 4,600 Zen 6 cores. Production shipments began in September; ramp "through the fourth quarter and into the first half of 2027". Initial systems at Meta, Microsoft, OpenAI, Oracle and Anthropic. AMD estimates 30% more tokens per dollar vendor estimate
Helion Announced
Open blueprint for agentic AI, Oct 12; partner configurations planned for November 2026. CPU-only rack with up to 32,768 cores (128 × 256), or an all-in-one rack with 16 MI350P GPUs (product page; one report said 32). Up to 102 Tbps networking, 21-inch ORv3. Eight launch partners; no pricing or rack power.
GPU D270 Sampling
Crescent Island named on Oct 12. No pricing, GA date or independent benchmarks. Intel claims one card serves ~300B-parameter models; Jon Peddie Research says that holds only on a 480 GB card or with FP4 weights, and puts dense decode near 5 tokens/s per stream. A Rubin GPU is about 13x its FP8.
Maia 300 Announced
Maia becomes "the primary inference system for Microsoft products", with multi-gigawatt deployments from 2027. Claims of 7x Maia 200 rack throughput and 40% more tokens per megawatt than leading GPU systems are vendor figures. Volume is reported at 300,000+ to 1 million chips. Azure customers cannot rent it; more detail at Ignite, Nov 17–20.
OpenAI Jalapeño Planned late 2026
OpenAI designed the system; Broadcom did physical design. 13.4 PFLOPS FP4, 232 GB at 15.4 TB/s, 550 W sustained and 700 W max, built for pods of 2,048 chips. OpenAI-cited benchmarks show up to 3.6x lower latency than GB300 vendor claim Small-volume deployment late 2026, ramping in 2027. Named a TIME Best Invention on Oct 15.
Broadcom AI semiconductor revenue, FY2026
Q4 is guidance. Hock Tan targets ~$115B of AI revenue in FY2027 and line of sight to $230B in FY2028 company targets Next report Dec 10.
Memory bandwidth per chip, hyperscaler and lab ASICs (TB/s)
Maia 300 leads the announced field on bandwidth per chip, but nobody outside Microsoft can use it. TPU and Trainium remain the custom silicon a buyer can actually rent.
Rentable today
TPU v7 Ironwood (192 GB) is GA on Google Cloud. AWS sells Trn2 and Trn3; Anthropic's agreement covers up to 5 GW across Trainium generations.
TorchNeuron In effect
Native PyTorch backend for Trainium2/3: public beta Oct 19, open-sourced under Apache 2.0 on Oct 20.
Values from one comparison table: Microsoft, IEEE Spectrum, AWS (via The Model Press). Status: Maia 300 announced; Jalapeño planned; Maia 200 in production; Trainium3 GA. Ironwood bandwidth not in material.
Capex keeps rising while the stress moves to neoclouds
Meta was the only one of the four big spenders to update its guidance in the window. Microsoft and Alphabet report on Oct 28, Amazon later.
Capex since 2023, from filings and guidance ($B)
Four-company 2026 guidance sums to $740–760B against $413B in 2025, roughly 80% growth. Against 2023, Amazon's plan is about 4.6x, Alphabet's top of range 6.3x and Meta's 5.2x.
Demand backs it
Alphabet (Jul 22): Cloud revenue +82%, a $514B Cloud backlog, continued supply constraints.
Investors are less forgiving
Alphabet fell 7% after its raise and turned cash-flow negative in Q2.
Source: MLQ compilation from SEC filings and guidance (2026 = top of range). Microsoft is not in the series.
Meta Q3 2026: spending through a litigation charge
Costs include a $9.89B charge for the August settlement with 52 attorneys general; without it operating income would have been $22.88B. Year-to-date capex of $85.83B implies Q4 capex of $49.2–59.2B. Reality Labs losses are being cut to about $15B, with the money shifted "primarily" to AI.
Neoclouds and their lenders
CoreWeave −15.46% on Oct 14 Market move
Gartner forecast that up to 80% of neoclouds may stop operating independently by 2030; CoreWeave closed at $65.78, Nebius −10.08%, Oracle −8.83%. CoreWeave cites a $104B backlog against $35.6B of debt principal; its 2032 bonds yielded about 15%.
Microsoft drops Loughton In effect
Anchor-tenant role at Nscale's 50 MW Essex site ended because grid power is not expected until the 2030s. The rest of the ~200,000 GB300 contract stands.
Anthropic–Nscale ~$45B In talks (reported)
460 MW of Vera Rubin capacity at the Monarch campus, West Virginia; would take Anthropic's Nscale commitments to about $90B.
Analyst projections
Goldman Sachs Estimate
US hyperscaler capex of $1.2tn in 2027 and $1.4tn in 2028. Hyperscalers need $308B a year of AI revenue to break even on 2026–27 capex; realized ROIC fell from 40% to 25%.
FactSet consensus, 2027 Estimate
$495bn in Sep 2025 to $1,131bn now (Apollo), about 2.3x.
UBS Estimate
Capex growth slowing from +76% in 2026 to +25% in 2027.
Credit conditions
AI-related debt issuance fell to about $23B in September, under half of August (36kr citing the FT). Neuberger (Oct 13) warns rising rates could slow AI capex.
Power, not chips, sets the build-out calendar
Read every "gigawatt campus" headline as two numbers: what draws power today and what is planned.
Largest AI campuses: operating vs planned IT power (MW)
No campus above 1 GW yet Estimate
Colossus 2 runs at about 62% of its plan, New Carlisle at 47%. Five campuses had been on track for 1 GW in 2026.
13.1 GW across 86 sites
About 17.0 GW with cooling overhead; average draw typically 60–80% of capacity (Sep 17 data). The single-site record has doubled about every 10 months.
Source: Epoch AI tracker, updated Oct 15, 2026. Estimates from satellite imagery and permits; about 80% within a factor of 1.4.
ERCOT: 92 GW asking for grid access
Deposits do not screen out speculation
At least $50M for a 1 GW connection against $30–40B to build the campus, or 0.13–0.17% of build cost.
Dedicated generation inside ERCOT
Pecos and Longhorn plan 2.7 GW and 1.3 GW of their own gas generation; Microsoft holds a 20-year Pecos agreement.
Raw queue vs buildable load
ERCOT's raw large-load queue was roughly 474 GW in August, about 90% data centers (Opportune); Modo's narrower 92 GW is used here.
Source: Modo Energy, Oct 21, 2026, unless noted.
The workaround: private gas and fuel cells
Venture Global Framework
GE Vernova and Baker Hughes frame agreements for up to 16 × 720 MW = 11,520 MW of power islands, supporting up to 10 GW of compute by end-2034. First order expected by end-2026; deliveries 2028–2034.
Google: Cheyenne and AEP Announced
Black Hills adds 564 MW of new gas plus about 2.1 GW of third-party resources (service from late 2027). AEP's $4B, 702 MW Bloom fuel-cell plant under a 20-year Google PPA, full operation by end-2028.
Policy friction
Texas paused data-center permits pending power and water audits; ERCOT's audit is due by Dec 10 Upcoming A PUCT rule requiring $50,000 per MW of security is in effect
Rack power: the 800 VDC chain arrives
Dell'Oro saw three shifts at OCP: the cold plate overtook the CDU as the focus, the 800 Vdc power sidecar "arrived in force", and solid-state transformers drew crowds well ahead of volume. MGX 1.2 supports racks up to 250 kW.
Delta in-row rack Showcased
Six 110 kW modules = 660 kW output, plus 480 kW of battery backup, for Vera Rubin NVL72; no ship date.
Still mostly demonstrators
LITEON: 95 kW DC/DC shelf with a 280 kW in-rack CDU. Rittal and Bosch: 800 VDC sidecar scalable to 2 MW, a development demonstrator.
Multi-megawatt CDUs at OCP ("table stakes")
Source: Dell'Oro; Delta, LITEON, Rittal and Bosch announcements.
Rental prices are rising and memory sets the premium
Two indices measure different things: Uvik's median of public on-demand list prices across 19 providers (captured Oct 7), and getdeploying's weekly median across more providers. They are not mixed below.
On-demand list price, $ per GPU-hour
Per hour, each newer generation costs more. The H200 is +36% over the H100 for +76% more memory, 23% cheaper per GB; the B200 is +95% and the B300 a further +16%.
Hyperscaler premium
H100 SXM: $10.53 at hyperscalers vs $3.95 at GPU clouds (2.7x). Provider quotes run from $3.20 to Azure's $12.29.
Per token, newer can be cheaper Estimate
1M Llama 2 70B output tokens: $0.18 on B200, $0.19 on B300, $0.36 on GB200 at median prices.
Source: Uvik, list prices in effect as of Oct 7, 2026. GB200 is a small sample (3 providers) with no hyperscaler split.
What moved in and just before the window
Nebius list increase In effect Oct 1
H100 $3.85 to $4.50 (+17%); H200 $4.50 to $5.40 (+20%).
AWS Capacity Blocks +~15% In effect Oct 7
On-demand prices unchanged. SemiAnalysis's one-year H100 contract index rose 38%, from $1.70 to $2.35, between Oct 2025 and Mar 2026.
Memory drives the B200 premium
In September a B200 cost 21% more per unit of compute than an H100, after being 18% cheaper in June: long-context inference needs KV-cache memory, and buyers now pay for it (SemiAnalysis via shattered.io).
List price vs market rate
H100 market indices sit about $1 below list medians: Silicon Data $2.81 (Oct 7), Ornn $2.86 (Oct 6), SemiAnalysis composite $2.82 (April). getdeploying's median discounts on $3.81: 1 year −28% (~$2.74), 3 years −49% (~$1.94), spot ~$2.21.
Own instead of rent?
B200 break-even at $30,000–40,000 per GPU is 3,663–4,884 GPU-hours, or 5.0–6.7 months at full use: a lower bound, since it counts only the GPU.
8×H100 for a month (5,840 GPU-hours)
AWS three-year commitment
$2.97 / GPU-hr
Cheapest GPU cloud (Hyperstack)
$3.20 / GPU-hr
GPU-cloud median
$3.95 / GPU-hr
AWS p5.48xlarge on-demand
$6.88 / GPU-hr
Hyperscaler median
$10.53 / GPU-hr
Selected option: $18,688 a month, 1.0x the cheapest GPU cloud and 1.1x an AWS three-year commitment.
Source: Uvik prices, Oct 7, 2026; 730 hours per month. Click an option to compare.
Tools, packaging and HBM cap supply into 2028
Wafers, lithography tools, packaging and memory cap how many accelerators can be built in the first place. Wafers are not the tightest link.
TSMC Q3: an AI quarter Reported Oct 15
Wafer revenue by node, Q3
79% from 7nm and below; gross margin 66.4%; full-year growth of about 44%. N2-family capacity should more than double in 2027.
Source: TSMC call, via MarketBeat.
ASML: supply short of 2027–28 demand Reported Oct 14
ASML plans about 30% more low-NA EUV capacity from about 65 machines, and the same for DUV immersion. CEO Christophe Fouquet: "30% in 2027, 30% in 2028 will not allow us to meet our customer demand." China's share of system sales fell to 23% from 42%.
Source: ASML investor release; Stock Titan.
CoWoS capacity, thousand wafers per month Estimate
TSMC does not publish CoWoS capacity. On media estimates it roughly doubles in 2026 (135 ÷ 70 ≈ 1.9x) and keeps growing through 2028; CoWoS is reported sold out through 2026.
GlobalFoundries interposers Announced Oct 8
About $2B over five years at Malta, New York; volume ramps in H1 2028, so TSMC keeps control of 2027 allocation.
OSATs on 2027–28 timelines
Amkor Arizona plan ~$12B; ASE panel line Q1 2027; SPIL Douliu Phase 1 in 2028; TSMC CoPoS volume 2H 2028 (media).
Source: Penchan (media and analyst estimates). Bars show the low end; the lighter cap shows the range where one is given.
HBM: sold out, repricing, and a wobble
Samsung
Shipping HBM4 since February 2026; 2026 volume effectively sold out. Measured HBM4 at 11.7 Gbps, about 3.0 TB/s per stack. Preliminary Q3 operating profit about 107.4 trillion won.
SK hynix
2026 DRAM and NAND output sold out; HBM4E mass production targeted for 2027. Shares fell 10.09% on Oct 14 on a report that some HBM4 back-end equipment orders may slip to 2027 (no company comment).
Micron
HBM4 in volume since March 2026. Much of its 2027 supply reportedly locked in at higher prices. A 36 GB stack moving from about $600 to about $1,300 is a projection estimate
Source: RCR Tech; TechTimes; tech-insider.org. Plan on 2027 HBM being allocated and priced higher.
Plan 2027 as a supply-constrained market
The chips announced in the window mostly reach volume in 2027, while the inputs that limit them are booked or sold out over the same horizon. The question has shifted from "which chip" to "which capacity can be delivered, powered and paid for, and when".
Buyers of compute
What changed
Rental prices rose for four straight weeks; Rubin is reservation-only until A5X GA; hyperscalers charge 1.7–2.7x GPU-cloud rates.
Implication
Lock in Blackwell or H100 capacity for 2026–27 now; compare cost per token, not per hour; treat three-year hyperscaler commitments as a real alternative.
Builders
What changed
No campus has crossed 1 GW of operating power; turbines are reserved to 2034; permits paused in Texas; 800 VDC and multi-MW CDUs are the new baseline.
Implication
Power procurement and permits decide the schedule; factory-built pods shorten only the last step.
Investors
What changed
Four-company capex of $740–760B for 2026; Meta's Q3 free cash flow negative; neocloud credit repriced; custom ASICs take inference share at the margin.
Implication
Watch the gap between long debt and short contracts at neoclouds; expect merchant-GPU exposure to shift toward training and enterprise.
Three judgements
Near-term supply is fixed; 2027 is where the variance lies
ASML plans about 30% more EUV and DUV immersion capacity in each of 2027 and 2028 and says even that falls short. CoWoS capacity is estimated to roughly double in 2026. HBM4 for 2026 is sold out and 2027 supply is being priced higher. Any 2027 plan that assumes falling GPU prices conflicts with all three.
The memory-heavy generation costs more per hour but can cost less per token
The B200's list median is 95% above the H100's per hour, yet at high utilization its cost per million Llama 2 70B tokens is an estimated $0.18. Utilization, not list price, decides which side of that line a team is on.
Custom silicon changes who captures inference margin, not whether Nvidia sells out
Microsoft keeps a mixed fleet and still trains on GPUs, and Maia is not rentable by Azure customers. The silicon outsiders can rent remains TPU and Trainium. For most readers, custom ASICs will show up as lower prices for Copilot or OpenAI services rather than as a hardware choice.
Key dates
Oct 1, 2026
Nebius H100 and H200 price increases In effect
Oct 14, 2026
ASML Q3 results; 2027–28 supply statement Reported
Oct 15, 2026
TSMC Q3 results and Q4 guidance Reported
Oct 21, 2026
Meta Q3 results; 2026 capex $135–145B In effect
Oct 28, 2026
Microsoft and Alphabet Q3 results Upcoming
Late Oct 2026
Samsung full Q3 results (segment detail) Upcoming
Nov 2026
Google Cloud A5X (Vera Rubin) general availability Upcoming
Nov 2026
AMD Helion partner configurations available Upcoming
Nov 17–20, 2026
Microsoft Ignite: more Maia 300 detail Upcoming
By Dec 10, 2026
ERCOT power audit under the Texas permit pause Upcoming
Dec 10, 2026
Broadcom fiscal Q4 results Upcoming
End-2026
Venture Global first firm turbine order expected Upcoming
Late 2026
OpenAI Jalapeño small-volume deployment Planned
Early 2027
First DSX POD deliveries by partners Upcoming
H1 2027
First Anthropic MI450 gigawatt begins Announced
2027
Maia 300 production deployment; HBM4E volume targets Upcoming
H1 2028
GlobalFoundries CoWoS interposer volume Upcoming
Signals that would change the picture
Hyperscaler Q3 calls (Oct 28 onward)
A cut to 2026–27 guidance removes the anchor for the $1.2tn 2027 analyst path; a further raise confirms Meta's direction.
Independent Rubin and Maia numbers
First GA A5X benchmarks and Ignite detail will show whether the 5–8x vendor ratios hold.
Neocloud funding
Whether Anthropic–Nscale signs, and whether AI-related debt issuance recovers from September's ~$23B.
HBM allocation
Whether SK hynix's reported equipment delay is confirmed, and how 2027 HBM4 contract prices settle.