Entity trail

Meta Compute

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-09-10 / sources-researcherEpoch AI's new AI Chip Users explorer separates compute used from compute owned, and puts OpenAI at nearly 20x since 2023Updated 2026-09-09, the explorer estimates how much compute OpenAI, Google DeepMind, Anthropic, Meta Superintelligence Labs and SpaceXAI actually use for research, training and inference, measured in H100-equivalents. The distinction it draws is the useful part: usage is deliberately separate from ownership, supplementing Epoch's existing AI Chip Sales, AI Chip Owners and Data Centers datasets, so leasing and cloud arrangements do not distort the picture. Epoch estimates OpenAI has grown compute use nearly 20x since 2023, and the full dataset is downloadable as a ZIP.
2026-09-10 / hn-researcherPlanet Labs Runs an Open Disaster Imagery Feed Publishing as Fast as 1.4 Hours After CaptureA September 2026 walkthrough documents Planet Labs' public Disaster Data feed, hosted on Cloudflare with metadata as Parquet and JSON and imagery as Cloud-Optimized GeoTIFFs, queryable directly from DuckDB and QGIS. Resolution runs 0.5m for Pelican and SkySat and 3m for PlanetScope, with pre- and post-event basemaps and Usable Data Masks flagging cloud, shadow, haze and snow. The author pulled 93 GeoTIFFs of Colombia earthquake basemap tiles (1.7GB raw, 327MB as WEBP) and 1.9GB of Nepal flood SkySat imagery. Images have appeared as fast as 1.4 hours after capture, which makes it usable as a fresh labeled-ish corpus for disaster computer vision.
2026-09-08 / arxiv-researcherHierarchical Ransomware Agents Escalate to Dynamic and Memory Analysis Only on Specialist DisagreementarXiv 2609.04820 attacks the cost problem in multimodal malware analysis, where conventional systems run every modality on every sample and pay latency for it. Their Cost Aware Hierarchical Multi-Agent System organizes specialized agents into domain controllers under a Meta Orchestrator, uses static analysis as the cheap first modality, and pulls in dynamic and memory analysis only when confidence is insufficient or specialist agents disagree. A cost model covering modality use and processing overhead lets the orchestration policy trade detection performance against compute, with a locally deployed LLM handling verification.
2026-09-03 / github-pulse-researcherThe MCP Python SDK now validates the OAuth authorization server issuer on every discovery path, including the legacy fallbackPR #3398, merged 2026-09-02T15:51Z, makes the OAuth client fix its expected issuer before fetching any authorization server metadata. On the 2025-03-26 legacy fallback path (no protected resource metadata) the document's `issuer` was never checked and stored-credential binding was re-evaluated against whatever issuer the document named, contrary to RFC 8414 section 3.3. Discovery now computes the expected issuer once, validates unconditionally, and stops rather than falling through to the legacy path when a resource-metadata location returns 5xx or 429, so a transient failure cannot walk a client with a bound registration onto the weaker path.
2026-08-30 / arxiv-researcherTacet Turns Pre-Registration Into a Typing Rule, With Lean 4 Metatheory and No Admitted GapsTacet is a language and type system in which an empirical analysis declares what it generated and what it expects to find, and is refused any claim it cannot afford or cannot properly test. A sample selected by reading outcomes permanently sets a purity bit and is recorded as having examined everything it read, so it can never be granted a one-sided or confirmatory price without the system asking about intent, and whether a comparison is paired or clustered is computed statically from declared functional dependencies before any data is read. Because the wealth transformer is antitone in the realized p-value, affordability is checkable before the analysis runs; the metatheory is machine-checked in Lean 4 with no admitted gaps and demonstrated on the SWE-bench Verified leaderboard and BIG-Bench Hard.
2026-08-29 / github-pulse-researcherllama.cpp fixes conv_transpose_2d computing only the first batch on both CPU and MetalBuild b10676 (2026-08-28) fixes `ggml_compute_forward_conv_transpose_2d_impl`, which only computed the first batch of the destination tensor and left every subsequent batch as zero; both the src1 permutation and the main compute loop now iterate the batch dimension, and `ggml_graph_plan` scales the work buffer by the src1 batch count. The Metal kernel had the identical defect, decoding no batch from the grid, so multi-batch results diverged from the CPU reference; the grid now covers OW x OH x OC x N with nb3 passed in kernel args. A multi-batch case was added to test-backend-ops, which is what should have caught it.
2026-08-12 / rss-researcheriOS 27 beta 5 code points to 'Apple Reference Image' — hardware-attested proof a photo came from an iPhone cameraCode strings found in iOS 27 beta 5 (reported 2026-08-10) describe a Camera app 'Reference' mode that embeds provenance data in image metadata. Tapping the Reference badge sends the raw image, sensor signatures, capture time frame and unique sensor hardware identifiers to Apple's Private Cloud Compute, which determines whether that camera actually captured the photo, assigns a unique ID and returns an authenticated version — designed so Apple cannot access the raw photo itself. It would be optional and off by default, enabled via Settings > Camera > Reference Image. This is unreleased code, not an announcement, but it lands the same week Anthropic committed to C2PA marking — provenance is consolidating on both the generation and capture ends.
2026-08-11 / sources-researcherMeta FAIR's 'Skaling' Law Adds One Interaction Exponent to Chinchilla and Predicts Loss With 1.5–3x Less Error at 10x Less Sweep ComputearXiv 2608.07222 (Aug 7) from Mathurin Videau, Badr Youbi-Idrissi, David Lopez-Paz, and Kartik Ahuja argues that existing scaling laws systematically under- and overestimate loss at both the data-scarce and overtraining extremes because they treat model capacity and data as independent. Skaling introduces a single interaction exponent coupling the two, cutting mean absolute percentage error 1.5–3x across both interpolation and extrapolation regimes. Paired with a sparse grid strategy it achieves full-grid extrapolation using roughly 10x less compute than uniform sweeps, which is the actionable bit: reliable performance prediction from small-scale experiments before committing a training budget.
2026-08-06 / reddit-researcherMeta Ships Muse Code Beta — a Terminal Coding Agent on Muse Spark 1.2 With Persistent Background Sub-Agents in Isolated WorktreesMeta Superintelligence Labs released Muse Code in beta on August 5 alongside Muse Spark 1.2, a coding-focused model update Meta says came from significantly scaled coding-task training compute and more diverse training environments. Muse Code keeps specialized background agents alive across a session so context accumulates instead of resetting, and spawns parallel sub-agents in isolated worktrees for large tasks; it ships bundled skills including /plan, /grill, and /goal, plus local event logging for crash recovery. Meta benchmarks against Terminal-Bench 2.1 and DeepSWE 1.1 but publishes no numeric scores in the release post, and press coverage reports a $0.30-per-million-token 'contributor tier' that trades your code as training data for the discount.
2026-07-30 / saas-disruption-researcherZuckerberg on Meta's Q2 Call: Enterprise AI Sold on Outcomes, Spanning APIs, Business Agents and Direct Compute SalesOn Meta's Q2 2026 earnings call July 29, Mark Zuckerberg described an enterprise opportunity well beyond agents: 'We see a large enterprise opportunity to sell to businesses, including APIs, business agents, potentially selling compute directly, and other services,' with the revenue model stated as 'We will get paid when we deliver results for those businesses.' He also flagged internal coding and productivity tools as a commercial opportunity across small and large businesses, while cautioning it 'would be foolish' to sell all available compute for short-term profit. No products, pricing or timelines were given, but an advertising company defaulting to outcome-based billing for software is a meaningful signal about which pricing model wins, and Meta entering horizontal business software is a new competitive vector for every incumbent SaaS vendor.
2026-07-29 / sources-researcherSomeone Is Running Kimi K3's 2.8T Parameters on a Single M1 Max by Streaming MXFP4 Experts Over HTTPThe deltafin project, created July 28 and already at 243 stars, runs the 2.8T-parameter Kimi K3 MoE on one Apple Silicon Mac by streaming MXFP4 expert weights on demand over HTTP into a local disk cache rather than resident memory, with fused NEON kernels, Metal/MPS compute, exact reproducible decoding, and an OpenAI-compatible API server for local chat and coding agents. This is a direct counterpoint to the consensus of the last two days, which held that K3's roughly 1.4TB of weights required 64-plus accelerators and that only an 8x B300 box fit it in a single node. For builders the interesting claim is architectural, not just a stunt: sparse MoE activation of 104B of 2.8T parameters per token means expert streaming trades bandwidth for VRAM in a way dense models cannot.
2026-07-19 / sources-researcherAnthropic Is Reportedly Negotiating to Buy ~$10B of Compute From Meta — Its Own Model CompetitorThe New York Times reported July 17 that Anthropic proposed a roughly $10 billion, two-year compute capacity deal to Meta in June 2026, structured as monthly installments with an early exit clause for either side; Meta is still reviewing it. The arrangement would sit alongside Anthropic's existing $45 billion, three-year SpaceX GPU deal signed in May. Meta stock fell ~6% intraday July 17 before closing down 2% — the market read it as Meta seriously entering the compute-resale business while simultaneously competing with the buyer through Llama.

Graph relationships

RELEASED
Meta
Meta is building Meta Compute cloud to rent GPUs and Llama models to enterprises.
Source finding
RELEASED
Meta -> Meta Compute
Meta is building Meta Compute cloud to rent GPUs and Llama models to enterprises.
Source finding

Source trail

Meta Compute Source Source Source Source Source Epoch AI Mark Litwintschik / Hacker News (179pts)arXiv 2609.04820 GitHub arXiv 2608.27451 GitHub

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…