LLM Inference Platform (class, 2022–present)

Class card for large language model inference providers as a machine type (OpenAI API, Anthropic API, Hugging Face Inference Endpoints, Mistral platform, Google AI Studio/Gemini, Together AI, Replicate, Groq, Cerebras, Fireworks.ai). Distinguishing features from the DM-instance counterpart machine:openai-foundation-model-lab-2015: this card is the LM-era abstraction — the class of API-mediated inference machines that inhabit the Cloud layer of Bratton's Stack without having built it. Individual labs are notable_instances; the class captures the shared operational grammar. Core operational grammar: (1) motor is absent — no extrinsic telos of Progress; LLM inference platforms are constituted by the feedback between user-demand signals and capability-capability deployment cycles without a directing Wissenschaft ideal; (2) substrate is inanimate + cognitive — trained model weights are inanimate tertiary-retention artifacts (Stiegler) that produce cognitive output at inference time; the class does not own the corporeal datacenter substrate (that is held by hyperscaler DM machines); (3) pace_layer is commerce, not infrastructure — pricing pressure, token cost, and model-release cadence are commerce-pace; (4) regime is chaotic [EXTRAP] — capability overhang, alignment uncertainty, and governance fragmentation make the class chaotic in the Cynefin sense. Cross-era position: heavy substrate_provision dependency on MM-era electrical grid (multi-GW inference load); substrate_provision from DM AWS cloud; parasitic_extraction against MM Humboldtian research university (training corpus + PhD talent without sustaining the system); parasitic_extraction against MM newspaper/broadcast info-substrate (LLMs trained on news archives without sustaining the publication-revenue model). Optional: zombie_dependency on MM industrial-era patent system (copyright law applies to training data; NYT v. OpenAI, Authors Guild v. OpenAI in progress). Compute concentration intensifies: AWS, Azure, GCP provide >95% of inference substrate; capture_resistance_index is LOW (~0.30 [EXTRAP]). Proletarianization signal (Stiegler): VERY HIGH (0.75 [EXTRAP]). The AGI narrative obscures the underlying ML-engineering competence; if no living competence re-internalizes transformer- training craft, inscribed weights persist but the lineage terminates. lineage_substrate: technical_memory (training datasets + model weights + API specs as Stiegler tertiary retention). All quantitative state-variable values are [EXTRAP]; framing is [CANON] per Wave-0 LM definitions and Wave-6 cross-era-coupling-typology §5 Stub 3 seed.