// PageInference — Fact sheet 03. Buyer-education page covering what inference
// is, why demand is outpacing supply, the structural reasons hyperscalers
// cannot close the gap quickly, and how Pyrra deploys distributed capacity to
// meet demand in weeks. Visually matches /specs and /financials.
const InferenceLinkCard = ({ onClick }) => (
{ e.preventDefault(); onClick(); }}
>
);
// ---------------------------------------------------------------------------
// Tiny composite: section title + pill tag, used over each fact table.
// ---------------------------------------------------------------------------
const TableSub = ({ title, pill }) => (
{title}
{pill ? {pill} : null}
);
const PageInference = ({ go }) => {
const trainingVsInference = [
{ title: "Purpose", body: "Training builds the model weights. Inference uses those weights to produce a response." },
{ title: "Frequency", body: "Training is episodic, measured in runs per quarter. Inference is continuous, measured in requests per second." },
{ title: "Workload shape", body: "Training is a long, synchronous batch job across thousands of GPUs. Inference is many small, latency-sensitive jobs across many smaller pools." },
{ title: "Locality sensitivity", body: "Training tolerates distance. Inference is round-trip-time sensitive and rewards being close to the user." },
{ title: "Cluster size", body: "Training favors single large clusters in one campus. Inference favors many smaller clusters distributed across metros." },
{ title: "Hardware fit", body: "Both run on Blackwell-class accelerators. Inference uses memory bandwidth and interconnect at the rack level rather than the campus level." },
{ title: "Spend curve over time", body: "Training spend is bursty and tied to model releases. Inference spend grows with adoption and compounds across every deployed product." },
{ title: "Industry sizing", body: "Industry analysts increasingly project inference spend will exceed training spend by a wide multiple over the second half of the decade, as agents and embedded models reach scale." },
];
const demandDrivers = [
{ title: "Foundation-model usage", body: "Aggregate token volume across major LLM providers has grown by orders of magnitude year over year since 2023, with no observable slowing in adoption." },
{ title: "Agent workloads", body: "A single autonomous agent run can consume tens to hundreds of times more inference per user task than a single chat turn, multiplying per-user demand sharply." },
{ title: "Embedded copilots", body: "Inference calls are now wired into IDEs, productivity suites, design tools, customer support flows, and operational dashboards. Each call adds to baseline load." },
{ title: "Reasoning-style models", body: "Modern reasoning techniques amplify the token count required for a single response by five to twenty times relative to a direct answer, raising compute per request." },
{ title: "Multimodal workloads", body: "Image, video, and voice generation and understanding each cost meaningfully more inference per request than text, and adoption of all three is climbing." },
{ title: "Latency expectations", body: "Product teams target sub-second responses for chat and sub-100-millisecond responses for embedded suggestions, which sets a floor on how distant the serving infrastructure can sit." },
{ title: "Enterprise commitments", body: "Multi-year inference offtake commitments from enterprise buyers are now common, locking in supply ahead of usage and tightening available capacity for spot buyers." },
{ title: "Sovereign demand", body: "National AI strategies in the US, Middle East, Europe, India, and Southeast Asia are funding inference capacity inside their own jurisdictions, removing supply from the global merchant market." },
];
const hyperscaleConstraints = [
{ title: "Grid interconnect queues", body: "Major US data center metros report multi-year waits for new utility interconnect capacity above the 100 MW threshold. The same pattern repeats in Ireland, the Nordics, and parts of Australia." },
{ title: "Permitting and entitlement", body: "Greenfield mega-campus builds typically take three to five years from site selection through commissioning, with permitting alone consuming twelve to twenty-four months in most jurisdictions." },
{ title: "Power generation lead time", body: "New gas, nuclear, or large-scale renewable generation tied to data center demand cannot be built faster than the underlying generation asset, which is measured in years." },
{ title: "Water availability", body: "Many hyperscale designs depend on evaporative cooling and lose access to favored locations when local water authorities cap or deny new draw permits." },
{ title: "Land and zoning", body: "Communities that previously welcomed data center campuses are imposing moratoria, design constraints, or tax-treatment changes that lengthen siting cycles." },
{ title: "Capex concentration", body: "A single hyperscale campus represents multiple billions of dollars of capex on a long-dated payback. Hyperscalers cannot diversify across many metros simultaneously without exceeding their internal hurdle rates." },
{ title: "Geographic concentration", body: "The supply that does exist concentrates in a small number of metros (Northern Virginia, Phoenix, Dublin, Singapore, and a few others). Buyers outside those metros pay a latency and routing premium even when capacity is technically available." },
{ title: "Sovereignty mismatch", body: "A workload requiring local jurisdiction cannot run inside a hyperscale region that sits in a different country, regardless of available capacity." },
{ title: "Refresh cadence vs build cadence", body: "Accelerator generations turn over roughly every eighteen months. Hyperscale construction takes multiples of that, meaning a campus broken ground today commissions onto silicon that is two generations behind the merchant frontier." },
];
const pyrraSpeed = [
{ title: "Unit size", body: "30 to 55 square feet of footprint, 25 to 200 kW of power draw. Sized to fit inside infrastructure that already exists in most commercial and institutional buildings." },
{ title: "No new substation required", body: "Each node lands inside the spare envelope of an existing commercial electrical panel. No new utility interconnect, no transmission build, no multi-year wait." },
{ title: "Cooling reuse", body: "Direct-to-chip liquid loops tie into existing chilled water, cooling tower, or dry cooler infrastructure already present at the host site." },
{ title: "Install duration", body: "5 to 7 calendar days for a P1 once MEP-ready. 10 to 14 calendar days for a P2." },
{ title: "Permitting envelope", body: "Installation falls inside the commercial mechanical permit category in most jurisdictions, which is measured in weeks rather than years." },
{ title: "Capex distribution", body: "Each node is independently owner-financed, which means Pyrra can stand up many nodes in parallel across many metros without concentrating capex on a single balance sheet." },
{ title: "Fleet operation", body: "Pyrra's network operations center carries every node under a uniform standard, so buyers see one network, one API, and one SLA across a globally distributed footprint." },
{ title: "Refresh path", body: "Because each node is a self-contained chassis, hardware refreshes happen at the unit level rather than the campus level, keeping the fleet close to the merchant frontier as new accelerator generations land." },
];
const compareRows = [
{ label: "Time from greenlight to live capacity", values: ["3 to 5 years", "Weeks once a host site qualifies"] },
{ label: "Unit of capacity", values: ["50 MW to 500 MW campus", "25 to 200 kW unit"] },
{ label: "Capex per unit", values: ["Hundreds of millions to billions, single balance sheet", "$1.5M to $2.8M per node, distributed across independent owners"] },
{ label: "New utility interconnect required", values: ["Yes, typically a multi-year process", "No, lands inside existing commercial electrical service"] },
{ label: "Permitting category", values: ["Greenfield industrial build, multi-stage entitlement", "Commercial mechanical equipment install"] },
{ label: "Geographic reach", values: ["A handful of campuses per region", "Every metro where a qualifying building exists"] },
{ label: "Locality to buyers", values: ["Routed across long-haul fiber from a distant region", "Inside the metro, often inside the building or campus serving the buyer"] },
{ label: "Refresh granularity", values: ["Whole campus, sized to the original power envelope", "Chassis-level swap on the fleet refresh cadence"] },
];
return (
Inference facts
The workload that defines the next decade of AI.
Inference is the running of a trained model against live requests: every chat reply, every
embedded suggestion, every agent action, every search summary. It is the workload that
converts AI capability into AI usage.
{/* ============================ HERO DIAGRAM ============================ */}
{/* ============================ 01 — TRAINING VS INFERENCE ============================ */}
01Training vs inference
One trains the model. The other runs it forever.
Buyers often conflate the two. They behave differently, scale differently, and need very
different infrastructure shapes.
The shape of the inference demand curve is no longer in doubt. The supply side is what
determines whether buyers can serve that demand at a price and latency the product needs.
Hyperscale data center construction was the right answer for the training era. The same
pattern stalls when applied to the inference era, for a list of reasons that have nothing
to do with effort or capital.
{/* ============================ 05 — HOW PYRRA DEPLOYS FAST ============================ */}
05How Pyrra meets demand quickly
Capacity that lands in weeks, not years.
Pyrra collapses the deployment cycle by reusing existing commercial mechanical and
electrical infrastructure and by miniaturizing the unit of capacity. The result is a
deployment cadence measured against building install windows rather than utility
construction cycles.
{/* ============================ 06 — SIDE BY SIDE ============================ */}
06Side by side
The build cycle, compared.
{/* ============================ 07 — THREE AUDIENCE CARDS ============================ */}
07Who this is for
);
};
// ---------------------------------------------------------------------------
// Hero diagram: a trained model fans out to many inference requests
// (chat, agent, search, embedded copilot, voice, vision).
// ---------------------------------------------------------------------------
const InferenceFanout = () => (