Vertically Integrated Compute

The standard taxonomy of AI compute has three tiers: hyperscalers (build and rent at scale), neoclouds (buy GPUs, rent them out), and colocation operators (rent shell and power). Tesla's Cortex 2 at Giga Texas is a fourth pattern — a company outside all three building its own AI-training datacenter for its own workloads, and pairing it with on-site power storage rather than renting either the compute or the electrons.

Cortex 2 is reported at 500MW at full buildout, targeted operational mid-2026, training Tesla's FSD neural networks. Its cooling design is the technically interesting part: two chiller plants with at least 12 fan units — roughly double Cortex 1's cooling capacity — with cooling equipment housed inside structures rather than sitting at ground level as in Cortex 1. That is a datacenter-architecture change driven by density, and it connects this concept to Rack Power Density. Construction pace is fast: structural beams delivered 2026-01-05 were installed within two days.

The strategic significance is the power question. Cortex 2 is reported as pairing GPU clusters with on-site Megapack storage, which makes it a live test of frontier debate #5 — who owns the power: utilities, IPPs, or the compute buyer building behind the meter. If a manufacturer with an energy-storage business can self-supply a 500MW AI campus, the "power is scarce" constraint is partly a procurement constraint rather than a purely physical one, and the set of companies that can escape the grid-interconnect queue is larger than the hyperscaler four. That is a hypothesis this KB cannot yet test — one analysis-tier source is not enough.

The compute figures are now first-party (Q2 2026 8-K), and the disclosure unit changed. Tesla's Q2 2026 print discloses AI Training Compute in megawatts: Cortex 1 >90 MW (Production), Cortex 2 >115 MW (Production) — with onsite Texas compute "more than doubled (in terms of MW) during the first half of 2026" [primary: 8-K Ex-99.1, 2026-07-22]. This supersedes the KB's prior treatment of the H100-equivalent counts as the current disclosure: as of Q2-2026 Tesla reports the table in MW, not H100e (Q1-2026 was H100e), and Cortex 2 advanced Early Ramp → Production. Current disclosed onsite compute is >205 MW combined — so the 500 MW figure is a full-buildout/site-envelope projection from the analysis blog, not the printed number. Capex is also now printed: Q2 capex $5.79B (+142% YoY), driving FCF negative $(1.1)B, self-funded by a $1.2B cash drawdown. Where the print and the blog disagree (units, MW vs H100e, 500 vs 205 MW), the print governs.

Weak sourcing on the rest, explicitly. The Cortex-specific detail below (500 MW full buildout, cooling architecture, Megapack unit counts, GPU H100e counts) still comes from a single analysis-tier blog, and the GPU/Megapack figures are summary-derived from web-search snippets — recorded as directionally useful and flagged so they are never published as fact without primary verification.

Key Claims

Added 2026-07-23 (Tesla Q2 2026 print — the compute figures are now first-party)

• [ACTUAL] Tesla AI Training Compute, disclosed in MW: Cortex 1 >90 MW (Production), Cortex 2 >115 MW (Production); >205 MW combined onsite Texas, "more than doubled" in H1 2026. Unit basis switched from H100-equivalents to MW; Cortex 2 advanced Early Ramp → Production. Evidence: strong (first-party 8-K) (TSLA Q2 2026)
• The 500 MW figure is a full-buildout/site-envelope projection, not the printed number — printed Cortex 2 is >115 MW in Production. Evidence: strong (print governs over the blog) (TSLA Q2 2026)
• [ACTUAL] Q2 capex $5.79B (+142% YoY) drove FCF negative $(1.1)B, self-funded by a $1.2B cash drawdown — the vertically-integrated builder shows the same "stopped self-funding" signature as the hyperscalers, but funded from its own cash, not capital markets. Evidence: strong (first-party 8-K) (TSLA Q2 2026)
• Deeper vertical integration underway: an in-house Austin semiconductor fab for Tesla's own logic/memory chips ("reliable long-term supply") — own compute, own power storage, and now own silicon. Evidence: strong (first-party 8-K) (TSLA Q2 2026)

Prior claims (blog-sourced)

• Tesla Cortex 2 (Giga Texas): 500MW at full buildout, targeted operational mid-2026 — a full-buildout target; the print discloses >115 MW in Production. Evidence: weak (single analysis source) (Tesla Cortex 2)
• Purpose is first-party workloads — training Tesla's FSD neural networks, not rented capacity. Evidence: weak (single analysis source) (Tesla Cortex 2)
• Cooling architecture changed between Cortex 1 and Cortex 2: two chiller plants, ≥12 fan units (≈2x Cortex 1's cooling capacity), with equipment housed inside structures rather than at ground level. Evidence: weak (single analysis source) (Tesla Cortex 2)
• Compute is paired with on-site Megapack storage — a behind-the-meter-adjacent model that partially routes around the grid-interconnect queue. Evidence: weak (single analysis source) (Tesla Cortex 2)
• A fourth category exists beyond hyperscaler / neocloud / colo: the vertically-integrated first-party builder. Evidence: weak (single instance, single source — a category inferred from one example) (Tesla Cortex 2)
• [UNVERIFIED — summary-derived] Cortex 2 listed as "Early Ramp" with >130,000 H100-equivalent GPUs installed (Q1 2026 shareholder deck, via secondary coverage); Cortex 1 >100,000 H100-equivalent. Evidence: weak (summary-derived from web-search snippets, not primary-verified) (Tesla Cortex 2)
• [UNVERIFIED — summary-derived] By mid-June 2026, Cortex 2 reported at >230,000 H100-equivalent GPUs installed. Evidence: weak (summary-derived, not primary-verified) (Tesla Cortex 2)
• [UNVERIFIED — summary-derived] Megapack buildout expanding from ~130 units toward ~250 units (~975 MWh on-site storage) per May 2026 permit filings. Evidence: weak (summary-derived, not primary-verified) (Tesla Cortex 2)

Benchmarks & Data

Printed 2026-07-23 (first-party 8-K):

• Tesla AI Training Compute (MW basis): Cortex 1 >90 MW (Production), Cortex 2 >115 MW (Production); >205 MW combined onsite Texas ("more than doubled" in H1 2026)
• Tesla Q2 capex $5.79B (+142% YoY); FCF $(1.1)B; self-funded by $1.2B cash drawdown
• In-house Austin semiconductor fab in progress (own logic/memory chips)

Blog-sourced (unverified unless noted):

• Cortex 2 power capacity: 500MW at full buildout (targeted operational mid-2026) — full-buildout envelope; print discloses >115 MW in Production
• Cortex 2 cooling: 2 chiller plants, ≥12 fan units (~2x Cortex 1)
• Construction pace: beams delivered 2026-01-05 installed within 2 days
• [SUPERSEDED by the MW disclosure] Cortex 2 installed GPUs: >130,000 H100-equiv (Q1 2026) → >230,000 H100-equiv (mid-Jun 2026); Cortex 1 >100,000 H100-equiv — Tesla no longer reports the table in H100e
• [UNVERIFIED] Megapack: ~130 → ~250 units (~975 MWh on-site storage)
• Cortex 2-specific capex: GAP — the print gives total Tesla capex ($5.79B), not a Cortex line
• Grid interconnect vs behind-the-meter split: GAP — not disclosed

Open Questions

• How much of Cortex 2's 500MW is actually behind the meter vs grid-supplied? The Megapack storage is reported, but storage is not generation — this distinction determines whether Tesla genuinely escapes the interconnect queue or merely buffers it.
• What did Cortex 2 cost, and what is its capex-per-MW versus a hyperscaler or neocloud build? No figure in any source.
• Is this pattern replicable by companies without an in-house energy-storage business, or is Tesla a special case?
• Does first-party compute earn a better or worse return than renting from a hyperscaler at these scales? The KB has no framework or data for this.
• Which other companies are building first-party AI datacenters at 100MW+ scale? xAI's Colossus 2 (500,000+ GPUs, Memphis) appears in the existing NVIDIA source and may belong to this category — the KB has not classified it.
• Does the vertically-integrated tier compete with neoclouds for the same GPU and packaging allocation (see Advanced Packaging Bottleneck)?

Related Concepts

• Power: The Binding Constraint — this concept is the strongest live test of frontier debate #5, "who owns the power"
• Rack Power Density — the Cortex 1 → Cortex 2 cooling redesign is a density response
• Hyperscaler Capex Cycle — first-party builders are capex that never appears in the big-four aggregate

Backlinks

Pages that reference this concept:

• Tesla

Changelog

• 2026-07-22 — Created from 1 source (tesla-cortex-2-500mw-gigatexas). Opens a fourth compute-ownership category (vertically-integrated first-party) against the existing hyperscaler/neocloud/colo taxonomy. Single weak source; GPU and Megapack figures flagged as summary-derived and unverified.
• 2026-07-23 — Folded in Tesla's Q2 2026 8-K: the AI Training Compute table is now first-party and switched units to MW (Cortex 1 >90 MW, Cortex 2 >115 MW, both Production; >205 MW combined) — superseding the H100e counts. Recorded that the 500 MW is a full-buildout envelope, not the printed figure; added the Q2 capex/negative-FCF print and the new in-house Austin fab (own silicon). The category's founding example is now primary-sourced on the compute side.