Compute Availability Factor (CAF): A Framework for Measuring the Physical Deliverability of Artificial Intelligence Compute: From nominal accelerator capacity to effective and infrastructure-available compute

Q: What is the Compute Availability Factor (CAF)?

The Compute Availability Factor (CAF) is an empirical measurement framework developed by Richie Adomako at Serpens Inc. (ResearchGate Publication ID 414150400) that quantifies the fraction of hardware-effective AI compute that can actually be delivered in the physical world under real electrical (E), thermal (T), network (N), and operational (O) infrastructure constraints: CAF(t) = E(t) × T(t) × N(t) × O(t).

Q: What is the relationship between the CAF website, CAF Watcher, and the CAF Research Paper?

CAF is the formal mathematical methodology and research specification ('Compute Availability Factor: A Framework for Measuring the Physical Deliverability of Artificial Intelligence Compute', Adomako, 2026) rooted in the thesis 'When the Building Becomes the Computer'. This site hosts both the formal specification and CAF Watcher—the operational observability platform, 3D Spatial Desktop OS, Cluster Physics Simulator, and Workload Constraint Investigator built on that methodology.

Q: How does CAF calculate Physically Deliverable Compute and the Deliverability Coverage Ratio (DCR)?

CAF first converts Nominal Compute (C_N, raw installed accelerators) into Hardware-Effective Compute (C_E = C_N × U_H). It then multiplies by CAF = E × T × N × O to derive Physically Deliverable Compute (C_D = C_N × U_H × CAF). The Deliverability Coverage Ratio (DCR = C_D / C_R) compares verified deliverable compute against roadmap required compute (C_R) using continuous interval bounds [Lower, Upper].

Q: How does CAF handle missing or unverified public infrastructure data?

CAF enforces the epistemic axiom that 'Unknown does not equal Unavailable'. Missing public evidence is never treated as zero capacity; instead, it is classified as UNKNOWN, widens the [Lower, Upper] availability interval, and reduces the separate Evidence Confidence score (Q).

Adomako, Richie

↳ CAF WATCHER // CONTINUOUS PHYSICAL AUDIT LEDGER

Cerebras Systems Grade A-

The Wafer-Scale Engine paradigm: bypassing chip packaging by turning an entire 300mm silicon wafer into a single computer · Category: Supercomputing & National HPC

Executive Physical Feasibility & Delivery Audit

Institutional Baseline: Evaluated under the Compute Availability Factor (CAF v1.0) specification by Richie Adomako (Serpens Inc., ResearchGate ID 414150400). CAF evaluates stated accelerator capabilities against empirical physical reality (interconnection queues, electrical substations, liquid cooling heat transfer, and optical network fabrics). Missing public evidence is classified as unknown—it is never treated as zero capacity.

CAF INTERVAL
[0.43 – 0.73]
DELIVERY GAP SCORE
32 pts
OPERATING POWER
0.1 GW
vs 0.1 GW announced
EVIDENCE CONFIDENCE
MEDIUM (65%)

Physical Coupling Pillars (E, T, N, O)

Power & Grid Substrate (E = [object Object]): 84/100. Delivers 23 kW per 15U chassis, requiring 415V three-phase electrical whips and custom power distribution units.
Thermal Dissipation & CDU Limits (T = [object Object]): 88/100. Engineered a monolithic cold plate that clamps directly against the 300mm silicon wafer with uniform micro-channel water distribution.
Cluster Topology & Optical Fabric (N = [object Object]): 86/100. Because all communication on the wafer happens via silicon interconnects at 214 Petabits/sec, intra-node networking bottlenecks disappear.
Silicon & Operational Availability (O = [object Object]): 94/100. By not cutting the wafer, Cerebras avoids the global CoWoS-L packaging queue that delays NVIDIA and AMD accelerators.