SILICON SIGNAL — Cloud · CPU · GPU Intelligence Desk

§ 01 · Lead

The Big Three Stories

What shifted in the cloud-silicon landscape this cycle, and why it reprices the competitive map.

▸ GPU · Platform Shift · Priority One

Nvidia's Vera Rubin goes into full production — and the buying cycle has already started without you.

Six chips shipping concurrently on a single platform. Nvidia's cadence is no longer "one or two new parts a year" — it is a yearly full-stack replacement, and the hyperscalers have voted.

Jensen Huang used his CES keynote to confirm that the Vera Rubin platform — Vera CPU, Rubin GPU, NVLink 6, ConnectX-9, BlueField-4, and Spectrum-X — is in full production, positioned as a 3.5× training and 5× inference uplift over Blackwell. The inference-cost story is the one to watch: Nvidia claims one-seventh the cost per token versus Blackwell, with GPU counts for MoE training down 75%.

Microsoft Azure and Google Cloud both committed to Rubin NVL72 in H2 2026. Azure says its Fairwater sites in Wisconsin and Atlanta were designed for Rubin's power/thermal envelope years ago. Meanwhile, hyperscaler backlogs keep ratcheting up: Amazon at $244B, Google at $240B. The 2–3 year GPU bottleneck Huang describes is the backdrop for every procurement conversation this year.

Rubin vs. Blackwell · Training

3.5×

▲ Per-GPU uplift

Inference Cost / Token

-86%

vs B200 platform

Nvidia FY26 DC Revenue

$197.3B

▲ 71.3% YoY

Custom Silicon · Azure

Maia 200 goes live and starts swinging.

Microsoft's second-gen accelerator is running workloads in Des Moines, powering GPT-5.2, Copilot, and the Superintelligence team's internal projects. Microsoft claims 3× FP4 over Trainium3, FP8 above TPU v7, and 30% better perf-per-dollar than Maia 100. On TSMC 3nm with 216 GB HBM3e.

GeekWire · Jan 26

CPU · AWS

Graviton5 is AWS's quiet margin machine.

Used by 90%+ of top-1000 AWS customers, Graviton5 claims up to 40% better price-performance than leading x86. Combined Trainium + Graviton run-rate is now $10B+ and growing triple digits — a structural challenge to Intel Xeon and AMD Epyc at cloud scale.

Amazon 8-K · FY2026

§ 02 · Landscape

The Accelerator Stack

Every chip that will matter to a cloud-scale AI buyer between now and year-end, at a glance.

Chip / Platform	Category	Peak Perf	Memory	Process	Status	Strategic Read
Rubin (VR200) Nvidia	Merchant GPU	~50 PFLOPS FP4 NVL72 rack · ~1.8 ExaFLOPS	288 GB HBM4 est., per chip	TSMC 3nm	H2 2026 Ramp	Default training fabric. Azure, GCP first in line.
GB300 (Blackwell Ultra) Nvidia	Merchant GPU	15 PFLOPS FP4 NVL72 · 92.1 ExaFLOPS	288 GB HBM3e per chip	TSMC 4NP	Shipping	The workhorse through 2026. Anchors Azure supercomputer.
Maia 200 Microsoft · Azure	Custom ASIC · Inference	10 PFLOPS FP4 5 PFLOPS FP8 · 750W TDP	216 GB HBM3e 7 TB/s · 272 MB SRAM	TSMC 3nm	Live in Prod	Microsoft's leverage vs Nvidia pricing. OpenAI workloads first.
TPU v7 (Ironwood) Google · GCP	Custom ASIC · Train + Inf	~4.6 PFLOPS FP8 Pod scale: ~9,216 chips	192 GB HBM3e per chip	TSMC 3nm	Shipping · 40%+ growth	The most mature hyperscaler ASIC. Price/perf lever for GCP.
Trainium3 Amazon · AWS	Custom ASIC · Training	~2× Trainium2 UltraServers link thousands	HBM3e Capacity not disclosed	3nm	Near-sold-out mid-2026	Anthropic's Project Rainier (500K+ chips) proves it at scale.
Instinct MI450 AMD	Merchant GPU	Next-gen CDNA Rack-scale platform	HBM4 est.	TSMC 3nm	2026 · Major customer landed	AMD's real Nvidia challenger. France national AI anchor deal.
WSE-3 (CS-3) Cerebras	Wafer-Scale · Train + Inf	125 PFLOPS FP8 1,000+ tok/s Llama-class	44 GB SRAM 21 PB/s on-chip BW	TSMC 5nm	Shipping · AWS Bedrock	$10B OpenAI deal. Now on AWS via disaggregated w/ Trainium.
LPU (GroqRack) Groq	Inference-Only ASIC	800–2,800 tok/s Llama 70B → Gemma 7B	230 MB SRAM No HBM/DRAM	GlobalFoundries 14nm	Shipping · GroqCloud	Deterministic latency. Real-time apps & voice AI winner.
Graviton5 Amazon · AWS	Custom ARM CPU	General cloud workloads 40% better price/perf vs x86	DDR5 Standard server mem	3nm-class	Shipping	Silent assassin of x86 share. 90%+ top-1000 AWS customers.
Core Ultra (Panther Lake) Intel	Client + Edge CPU	Hybrid P+E + NPU 18A foundational part	LPDDR5X —	Intel 18A	CES 2026 Launch	Intel's credibility test. Cloud impact mostly indirect via AI PCs.

§ 03 · Edge · Physical AI

The Robot Brain Stack

The CPUs and AI processors going into humanoids, robots, and autonomous vehicles — where inference moves on-device.

Chip / Platform	Form Factor	AI Compute	Memory	Power	Status	Strategic Read
Jetson AGX Thor Nvidia · Blackwell	Humanoid · Industrial	2,070 TFLOPS FP4 7.5× Orin · 3.1× CPU	128 GB LPDDR5X 14-core Arm Neoverse-V3AE	40–130 W	GA · Jan 2026	The de facto platform. Boston Dynamics, NEURA, Agility all adopted.
Jetson T4000 Nvidia · Blackwell	Mid-tier Robotics	1,200 TFLOPS Cost-effective SKU	64 GB LPDDR5X	40–70 W	Available · CES '26	Volume play — AMRs, delivery bots, smart tractors.
DRIVE Thor Nvidia · Blackwell	Autonomous Vehicle	~2,000 TFLOPS Unified world-model inference	Integrated SoC Sensor + AI fabric	~130 W	In Production Vehicles	Aurora, Waabi, BMW iX3 (Level 4). Ecosystem > silicon.
Dragonwing IQ-X Qualcomm	Humanoid · Service Robot	VLA + VLM optimized Snapdragon heritage	Integrated LPDDR —	< Jetson class	CES '26 · Figure partner	Power/efficiency play vs Jetson. Figure, KUKA, Advantech signed on.
Snapdragon Ride Pilot Qualcomm	ADAS · Autonomous Drive	Flex compute tiers Up to L4	— Auto-grade	—	BMW iX3 '26	Qualcomm's auto beachhead. Main Nvidia DRIVE counterweight.
Tesla AI5 (HW5) Tesla	FSD · Optimus · Vehicle	~H100 class 10× AI4 compute	~9× AI4 memory In-house design	Auto-grade	Taped out · H2 2026	Optimus first, not cars. Dual-source TSMC AZ + Samsung TX.
Dojo 3 (AI6) Tesla	Training Supercomputer	~2× AI5 Neural net training	— Server class	Data center	Tape-out Dec '26	Dojo returns. Trains the models AI5 deploys. Samsung fab.
Hailo-15 / Hailo-10 Hailo	Low-power Edge AI	Up to 40 TOPS Vision-centric	Integrated Host-paired	< 5 W typical	Shipping	Cameras, smart sensors, low-cost AMRs. Below Jetson's power floor.

▸ Customer Adoption · Who Ships What

Robotics & AV customers by silicon platform

Design-win concentration is the leading indicator of platform lock-in — follow the robots, not the roadmaps.

Jetson Thor · T4000

Humanoids & Industrial Robotics

Boston Dynamics · Atlas humanoid

Agility Robotics · Digit Gen 6

NEURA Robotics · Gen 3 humanoid

Franka · Humanoid · GR00T workflows

Caterpillar · LG · Industrial · appliances

Locus · Fetch · AMRs on Orin / T4000

◆ DOMINANT · Ecosystem moat

DRIVE Thor · AGX

Autonomous Vehicles

Aurora · L4 autonomous trucking

Waabi · Autonomous trucks

Nuro · Last-mile delivery

Zoox · Mercedes · Robotaxi · DRIVE

Volvo · JLR · OEM partners

Cruise · wound down (GM, 2024)

◆ STRONG · Default in L4 trucking

Dragonwing · Ride

Qualcomm Ecosystem

Figure AI · Humanoid · OpenAI-backed

KUKA Robotics · Industrial arms

Booster · Humanoid

Advantech · APLUX · Edge compute OEMs

BMW · iX3 Ride Pilot L2+

Robotec.ai · VinMotion

◆ EMERGING · Figure is the headline

In-House Silicon

Vertically Integrated Stacks

Tesla · AI5 · FSD + Optimus

Waymo · Custom silicon · 6th gen

Mobileye · EyeQ family · OEM ADAS

Hailo · Merchant edge ASIC

Rationale: volume, latency, IP control

◆ CONTRARIAN · Scale-dependent bet

▸ THE PATTERN Nvidia owns the reference design for humanoids and L4 trucking — Boston Dynamics, Agility, Aurora and Waabi all ship on Jetson or DRIVE. Qualcomm's early design wins (Figure, KUKA, Booster) cluster around Series-A/B startups optimizing for power and BOM cost; Figure is the headline, KUKA is the volume. The vertically-integrated camp — Tesla, Waymo, Mobileye — pays the CapEx to own their silicon because their deployment volumes (or ADAS royalty streams) justify it. Most new entrants will pick from the Nvidia or Qualcomm stack; building your own is a business-model decision, not a technology one.

▸ Key Insight · Stack Bifurcates

"Android of robotics" vs vertical integration

Nvidia is betting Jetson + Isaac + GR00T becomes the open standard. Tesla is betting custom silicon + in-house everything wins on volume. Both can be right.

▸ The Qualcomm Angle

Arduino acquisition + Dragonwing

Qualcomm's edge: power efficiency heritage from mobile. Partnerships with Figure, KUKA, Advantech suggest genuine Jetson alternative — but no Isaac-equivalent yet.

▸ What Matters Most

The foundation model, not the flop

GR00T N1.6, Cosmos Reason 2, Isaac Lab-Arena — robotics is software-defined. Pick silicon with the software stack, not the spec sheet.

§ 04 · The Money

Who's Buying, How Much

Capex is the leading indicator for every chip roadmap. In 2026, the top five will commit ~$600B.

▸ Hyperscaler Capex · 2026 Guide

$602B, three-quarters of it AI

Projected 2026 capital expenditures by company, with YoY growth. Source: company guidance, Dell'Oro, Introl.

Amazon AWS Heavy

$200B▲ 53%

Google GCP + Search

$180B▲ 37%

Microsoft Azure Core

~$160B▲ ~50%

Meta AI Infra

$115–135B▲ ~80%

Oracle OCI

~$70B▲ Accelerating

AI Share of Capex

~75%

$450B of the $602B

Capex / Revenue

45–57%

Utility-like intensity

2025 Debt Raised

$108B

$1.5T projected

▸ Accelerator Market · 2026E

Nvidia keeps training — custom ASICs take inference

Share of AI accelerator spend by type. Custom silicon CAGR to 2033: ~44.6% vs merchant GPU ~16.1%.

Nvidia GPUs Merchant silicon

~72%

Google TPU Ironwood v7

~10%

AWS Trainium Gen 2 + 3

~7%

AMD Instinct MI325X / MI450

~5%

MSFT Maia + Meta MTIA

~4%

Other Cerebras, Groq, SN

~2%

§ 05 · Feed

The Wire

Live from our tracked sources — rebuilt on every run of fetch_feeds.py.

    ● Live Feed
    Generated: — · Items: —
  

JavaScript required for live feed.

The news feed is rendered from embedded JSON at page load. Enable JavaScript to see the latest items from our tracked sources.

§ 06 · Strategy

The PM's Read

Three decisions your product and strategy teams should be pressure-testing this week.

▸ Desk analysis · For internal planning

The cloud-silicon map has already repriced. Here's what matters for our roadmap.

If we're planning anything beyond a 6-month horizon that touches AI compute, cloud pricing, or memory availability, these three pivots frame the conversation.

Design for multi-silicon from day one

Nvidia still owns training, but inference is splintering fast — Maia 200, Trainium3, TPU v7 all claim parity or better on perf/dollar within their home clouds. Assume our workloads will run on at least two accelerator families in 18 months. Abstract the runtime now; the CUDA lock-in story weakens every quarter as vLLM, ROCm, and custom compilers mature.
Memory, not compute, is the binding constraint

DRAM up 125%, NAND up 234% in 2026 per Gartner. Any product decision that assumes stable memory pricing through 2027 is wrong. Renegotiate long-term cloud commits before Q3; avoid locking in pricing terms past 2027 without inflation adjustments. Memflation is PM-relevant — it'll show up in our COGS before it shows up in our infra bill.
Neoclouds — both threat and option

Meta's $27B Nebius deal is the permission structure. Even hyperscalers are capacity-constrained enough to rent. CoreWeave, Nebius, IREN, Lambda give us leverage in any pricing negotiation with the Big 3, and optionality on Rubin allocation. Brief CFO on neocloud economics; it's the cheapest hedge against hyperscaler lock-in we have right now.

§ 07 · Calendar

On the Radar

The four near-term events with the biggest potential to move our planning assumptions.

◦ May 5

AMD Q1 '26 earnings

MI450 pipeline commentary, data-center AI share, ROCm adoption metrics. Key question: is the French gov deal a one-off or the start of a pattern?

◦ Q2 '26

Nvidia Rubin VR200 first shipments

Azure and GCP allocation confirmation. Rubin NVL72 rack economics. Whether the 7× inference cost claim holds on real workloads.

◦ Mid-2026

Trainium3 capacity fully committed

AWS guided nearly all supply committed by mid-year. When it tips over, pricing power shifts materially in AWS's favor. Watch for Trainium4 guidance.

◦ Jul 1

Section 232 tariff review

US Commerce required to update POTUS on chip tariffs affecting data-center builds. Changes here ripple into every capex model downstream.