Colossus 2 vs Stargate vs Hyperion vs AWS — what's the difference?

Different ownership models and customer mixes. Colossus 2 (SpaceX, near Memphis) is a third-party AI infrastructure provider running Nvidia GB200/GB300 chips and currently serving Anthropic ($1.25B/mo), Google ($920M/mo), and Reflection AI ($150M/mo) — total $2.3B+/month. Stargate is OpenAI's joint venture with Oracle and SoftBank — primarily for OpenAI's own workloads. Meta Hyperion is Meta's owned infrastructure for Llama and product workloads. AWS is the broadest commercial provider with multi-tenant AI capacity and integrated services. Colossus 2 is the surprise of 2026 — going from internal training cluster to top-tier external provider in roughly two months.

How much AI compute is actually being built in 2026?

Combined committed spend across the largest 2026 AI data center programs runs into hundreds of billions through 2029. Colossus 2 alone is at $97.4B+ committed across three customers. OpenAI Stargate plans are reportedly in the $500B+ range over multiple years. Meta Hyperion has billions in 2026 capex alone. Google has $175-185B in guided AI infrastructure capex. Anthropic just signed 1 gigawatt of US data center leases. The grid infrastructure question is no longer theoretical — it's an active operational constraint on the AI infrastructure buildout. Power transmission and substation capacity are the practical bottleneck in mid-2026, not chip supply.

Is grid capacity actually limiting AI buildout?

Yes, in specific geographies. The most-cited constraint is the US Southeast and the PJM (mid-Atlantic) interconnection queue — projects requiring large new substations are facing multi-year wait times for interconnection studies and approvals. The Memphis area (where Colossus 2 sits) and Northern Virginia (the data center capital) have been hit hardest. Workarounds being used in 2026: behind-the-meter generation (gas turbines on-site), existing-facility expansion (avoiding new interconnection), and geographic spread to grids with capacity (Texas, Oklahoma, parts of the Midwest). Anthropic's 1 gigawatt of US lease commitments includes deliberately diversified geography for this reason.

Which AI infrastructure provider should an enterprise pick?

Depends on workload type and existing relationships. For training large models: typically not a choice — you go where you have a contract (Anthropic uses Colossus 2 + other; OpenAI uses Stargate + Azure). For inference at scale: AWS, Azure, and Google Cloud are the realistic choices for most enterprises because they integrate with existing services. For enterprise AI products: stick with your hyperscaler. The Colossus 2 / Stargate / Hyperion distinction matters most for frontier labs themselves and for understanding the macro picture. Reflection AI's deal opens the possibility that Colossus 2 starts serving more open-source-frontier labs in 2027.

Quick Answer

Colossus 2 vs Stargate vs Hyperion vs AWS — AI Data Centers (Jun 2026)

Published: June 23, 2026

Colossus 2 vs Stargate vs Hyperion vs AWS: The 2026 AI Data Center Map

Four large AI infrastructure programs define mid-2026 capacity: SpaceX Colossus 2 (just hit $2.3B/month committed), OpenAI Stargate (with Oracle and SoftBank), Meta Hyperion, and AWS’s multi-tenant AI capacity. Plus Google’s $175-185B guided capex. Plus Anthropic’s 1 gigawatt of US data center leases. Here’s how they compare, what each is good for, and where the real bottlenecks are.

Last verified: June 23, 2026.

TL;DR

Program	Owner	Customer model	Headline scale
Colossus 2	SpaceX	Third-party rental	$2.3B+/mo committed ($97B+ to 2029)
Stargate	OpenAI / Oracle / SoftBank	OpenAI-anchor	$500B+ multi-year reported plans
Hyperion	Meta	Meta-internal	Tens of billions 2026 capex
AWS AI infra	Amazon	Multi-tenant commercial	Integrated with broader AWS spend
Google AI capex	Google	Google-internal + GCP	$175-185B guided
Anthropic leases	Anthropic (mixed sites)	Anthropic-internal	1 GW US leases

Combined committed spend across the named programs runs into hundreds of billions through 2029. The bottleneck for further growth in mid-2026 isn’t capital — it’s grid capacity.

Colossus 2 — the surprise of 2026

SpaceX’s Memphis-area facility went from internal training cluster to top-tier third-party AI infrastructure provider in roughly two months. Current customer mix:

Anthropic: $1.25B/month → ~$52.5B through 2029 → 220,000 Nvidia GPUs
Google: $920M/month → ~$38.6B → 110,000 GPU equivalents
Reflection AI: $150M/month → ~$6.3B → GB300 chips (June 22, 2026 announcement)

All three contracts include a 90-day exit clause after the first three months — meaning ~$7B is contractually firm, and the rest is rolling quarterly customer decisions.

Strategic position: SpaceX has signed over $80 billion in committed compute revenue in roughly two months, positioning itself alongside AWS / Azure / Google Cloud as a major AI infrastructure provider — despite having no prior cloud business. The Colossus 2 GB200/GB300 stack is competitive at the chip level.

Risk: SPCX fell 10% on June 22 because the market started pricing in the 90-day-exit reality. Compute-as-a-service margins are lower than AI-product margins. SpaceX is becoming an infrastructure company on top of being a space company and an AI company.

Stargate — OpenAI’s mega-program

The OpenAI / Oracle / SoftBank joint venture. Plans publicly reported in the $500B+ range over multiple years. Stargate’s structure is different from Colossus 2: it’s anchored by OpenAI’s own training and inference workloads rather than third-party rental. Oracle Cloud Infrastructure provides much of the operational backbone; SoftBank provides capital.

Strategic position: Stargate is OpenAI’s bet that vertical integration of compute matters more than buying from third parties at scale. The GPT-5.5 / GPT-5.6 / future-model training runs depend on Stargate capacity coming online on schedule.

Risk: Grid interconnection delays affect Stargate sites as much as anyone else. The “build vs rent” calculus for OpenAI vs Anthropic has been a multi-year debate; Anthropic chose rent (Colossus, plus its own leases), OpenAI chose build-via-JV (Stargate).

Hyperion — Meta’s internal infrastructure

Meta’s AI infrastructure program — Hyperion — supports product workloads (Instagram / Facebook ranking, WhatsApp AI features, Ray-Ban glasses, Meta AI assistant) and the now-pivoting Llama / Muse / Spark research program.

Strategic position: With Meta unwinding the Llama open-source frontier program (announced June 2026), Hyperion’s role shifts toward product-AI and away from frontier research compute. The capex remains real but the strategic urgency has shifted.

Risk: The Meta AI product story is competitive but not dominant. If Meta’s AI-product roadmap doesn’t translate into clear revenue uplift, Hyperion’s capex justification weakens.

AWS AI infrastructure — the multi-tenant baseline

AWS remains the broadest commercial AI infrastructure provider. Trainium chips (Amazon’s custom AI silicon) and Nvidia-based instances both available. Integrated with the rest of the AWS service stack — RDS, S3, SageMaker, Bedrock — which matters enormously for enterprise adoption.

Strategic position: AWS doesn’t have to win the frontier lab business to win the AI infrastructure category. Most enterprise AI workloads are inference and lighter training, not frontier model pretraining. Bedrock (multi-model API) lets AWS sell access to Claude, Llama, Mistral, and other models without owning the underlying training.

Risk: Less. AWS is the safest tier of the major AI infrastructure programs because the business mix is already validated.

Google and Anthropic — the other big numbers

Google’s $175-185B guided AI infrastructure capex is for Google’s own use — TPU v6 and v7 buildout, plus Nvidia capacity (including the Colossus 2 contract). The numbers are real and material to AI infrastructure pricing dynamics.

Anthropic’s 1 gigawatt of US data center leases (announced earlier in 2026) is a hedge against single-provider concentration risk. Anthropic uses Colossus 2 for SpaceX-supplied capacity but is also leasing geographically diversified data center space directly — partly for redundancy, partly for grid-capacity hedging.

The actual bottleneck: grid capacity

The most important point in the 2026 AI infrastructure picture: chip supply is no longer the binding constraint. Grid interconnection is.

The PJM interconnection queue (mid-Atlantic) has multi-year wait times for projects requiring new substations
US Southeast grids are stressed by the data center boom
Workarounds in use: behind-the-meter gas turbines, existing-facility expansion, geographic spread to less-stressed grids (Texas, Oklahoma, parts of the Midwest)
Anthropic’s 1 GW lease portfolio is deliberately geographically diversified for this reason

This is the structural change that justifies the Colossus 2 / Stargate / Hyperion multi-billion-dollar competitive positions. Whoever has substations and interconnection already has a moat. Whoever is in the queue is at the mercy of regulators.

Implications for the rest of the market

For frontier labs: Compute access is becoming a binding constraint on which labs can stay frontier. Reflection AI’s $6.3B Colossus 2 deal is the entry ticket to staying in the game. Smaller labs without comparable compute commitments will fall behind on capability.

For enterprise buyers: The major AI infrastructure providers are differentiating less on raw chip count and more on regulatory/grid moats. AWS, Azure, GCP integration with existing enterprise services remains the dominant decision factor.

For investors: The infrastructure programs are real moats. The SPCX 10% drop is a margin correction, not a thesis-breaker. OpenAI’s Stargate execution risk is real but manageable. AWS / Azure / GCP AI businesses are durable.

Sources

AIToolsRecap, “AI News June 23 2026”
CNBC, MLQ News, Yahoo Finance on SpaceX / Reflection AI deal, June 22, 2026
BuildFastWithAI on Anthropic 1 GW lease commitment and Google capex
Public reporting on Stargate JV structure (Oracle, SoftBank)
andrew.ooo prior coverage on Colossus 1, Stargate, Hyperion, and SPCX IPO

Verified June 23, 2026.