Hyperscalers

Google Cloud revenue rose 82% YoY to $24.8B in Q2 2026. Capex guide lifted to $195–205B; TPU v7 is the largest non-merchant accelerator estate.

AWS grew 37% in Q2 2026 to a $169B annualised run-rate — its fastest in 18 quarters. Capex guidance was raised to $220B in July.

Guided to roughly $190B of CY26 capex, of which about $25B is attributed to higher component pricing — memory inflation showing up directly in a hyperscaler budget.

Highest backlog-to-revenue ratio in the group and now a serial debt issuer — RPO is dominated by a handful of AI tenants.

Three-year AI infrastructure program; in-house Pingtouge silicon paired with domestic DRAM.

Largest Chinese buyer; export controls push it toward domestic HBM and CXMT DRAM.

Slower capex ramp than peers; management has repeatedly flagged accelerator supply limits.

SpaceX absorbed xAI in February 2026 at a ~$250B mark; the combined entity is valued near $1.25T. xAI contributed ~$818M of SpaceX's ~$4.69B Q1 2026 revenue, and Colossus capex sits on SpaceX's own balance sheet.

Capex guide $130–145B with no cloud revenue offset — the purest memory-demand line in the group. Q2 operating margin fell to 31% from 43% a year ago.

Smallest datacenter footprint of the group, but the largest mobile LPDDR buyer by units.

Q1 2026 revenue $2.08B (+112% YoY) with a $99.4B contracted backlog and 3.5GW of total contracted power, but a $740M net loss — the backlog is debt-funded and rate-sensitive.

Inference-led pivot after the Eigen AI and Clarifai acquisitions; European power contracts insulate it from US grid queues.

Targeting $3.4B ARR by end-CY26 off a 140k GPU expansion, with >4.5GW of secured grid-connected power after the 1.6GW Oklahoma campus. Owns its power, the scarcest input after HBM.

Vertically integrated power plus compute; anchor build partner on very large single-tenant campuses.

Hit ~$760M annualised at end-2025 (+79% YoY) and was valued at $5.9B in its $1.5B Series E; killed the on-prem hardware line and now sells to frontier labs and hyperscalers.

Hosting-led economics: lower margin than pure GPU cloud, but power-secured and contracted long.

Nordic hydro-powered capacity; sovereign-AI contracts give it unusually long-dated backlog.

Inference-weighted mix; sensitive to token-price deflation rather than GPU rental rates.

Aggregator model with asset-light exposure; books frontier-lab training clusters.

Broad edge footprint, smaller cluster sizes — buys mid-tier accelerators and far more DDR5 per GPU.

Run-rate revenue crossed $47B in May 2026 and it raised $65B at a $965B post-money — now the largest pure-play lab by both revenue and valuation. Reportedly paying SpaceX ~$1.25B/month for xAI-cluster capacity.

~$25B annualised as of February 2026 with a 33% gross margin; inference costs projected at $14.1B and cash burn near $27B for 2026. Filed confidentially for an IPO in June 2026.

Absorbed by SpaceX in February 2026 at a $250B mark; the combined group is valued near $1.25T. Standalone AI revenue is still the smallest of the frontier three, but Colossus capex is self-funded.

Efficiency-first training recipes; the clearest downside risk to per-token memory demand.

European sovereign-AI anchor; smaller models pull DDR5 inference capacity more than HBM.

Kimi long-context models are KV-cache heavy — unusually high DRAM per served token.

State-linked deployments; constrained to domestic accelerators and CXMT-supplied DRAM.

On-prem and private-deployment focus; sells into customers who buy their own memory.

Early revenue, very large reserved training footprint relative to headcount.

Pre-revenue, but reserved compute at a scale that shows up in supplier order books.