Trusted access
Open weights on owned infrastructure, with tenant boundaries wired into the cache key.
Owned inference infrastructure
CueCode private inference cloud
Frontier-class open coding models on memory-dense Apple Silicon pods, sold per engineer seat behind an OpenAI-compatible endpoint.
~75%
lower seat cost than frontier power-user API spend
~5
floor-bounded decode streams per four-node pod
$110K
pilot pod to validate the load-bearing throughput number
The product
Predictable private inference for teams that live inside agents.
Engineering teams are already mixing Cursor, Copilot, Claude, GPT, Gemini, and internal agents. The result is expensive, hard to govern, and increasingly tied to metered usage. CueCode replaces that chaos with owned capacity: a per-seat line item, auditable usage, and private model serving on hardware with known provenance.
The service runs open coding models on four-node Apple Silicon pods. Customers keep the familiar API surface while CueCode owns the serving layer, scheduler, cache placement, metering, and isolation.
Predictable spend
Seats price concurrency instead of surprise token bills from heavy agentic work.
Control plane
SSO, budgets, audit logs, per-seat metering, and endpoint compatibility.
Dedicated paths
Pooled, reserved, and dedicated tiers map isolation and concurrency to price.
Why now
Open coding models are good enough. The access layer is not.
AI coding spend is fragmenting
Teams are buying multiple copilots, chat tools, and agents at once. Heavy users create the most value and the least predictable bills.
Native open APIs are not enough
The best open coding models are attractive, but enterprise buyers need provenance, governance, stable capacity, and clear jurisdiction.
Owned capacity changes the unit
A private pod turns metered token surprise into a seat contract: concurrency, isolation, auditability, and a fixed line item.
Why it works
Memory bandwidth, idle gaps, and a batcher built for coding agents.
MoE inference is memory-bound
The active weights are read every token. Apple Silicon unified memory gives the pod dense, low-power bandwidth where the workload needs it.
Agents are bursty
Coding agents decode, call tools, wait, and resume. CueCode backfills the idle gaps so one pod can serve many engineers without wasting cycles.
Aggregate tokens are the unit
One forward pass advances every admitted stream by one token. The business sells the aggregate capacity above the 12 tok/s floor.
Flow test
Good tok/s is the difference between waiting and staying in the problem.
Agentic coding is not a paragraph in chat. It is code, diffs, logs, traces, and tool output arriving while the engineer is still reasoning. Once the stream falls below the floor, attention breaks.
- Retention floor
- 12 tok/s
- CueCode target
- 60 tok/s
- Sample completion
- 4.6s
Unit model
Move the assumptions. Watch the pod economics move with them.
5.1
engineers per pod at this workload
- Single-stream speed
- 15.7 tok/s
- Aggregate output
- 55 tok/s
- Decode capacity
- 4.6 streams
- Capex payback
- 1.2 years
Model formula: single = 28 / (1 + context / 64K); aggregate = single x batch; capacity = aggregate / 12.
Seat shape
One pod model, three ways to buy the concurrency it creates.
Pooled
$1,500/seat-mo
Shared capacity for light and mid-market teams that want predictable spend without dedicated hardware.
Reserved
$2,500+/seat-mo
Guaranteed stream headroom for engineers running multiple agents, branch explorations, and long test-fix loops.
Dedicated
Custom/pod
Hardware-isolated pods for sensitive codebases, regulated teams, and the heaviest orchestration workloads.
The moat
The scheduler decides which streams occupy the scarce decode slots.
S1
S2
S3
S4
S5
Mac 01
L 1-20
L 1-20
Mac 02
L 21-40
L 21-40
Mac 03
L 41-60
L 41-60
Mac 04
L 61-80
L 61-80
waiting streams
Q1 Q2 Q3
Admission controller
fairness, floor, seat headroom
floor
residency
entitlement
independence
KV placement
resident, prefetch, tier to SSD
admitted batch
S1 S2 S3 S4 S5Step 1
Retire finished streams
Finished, stopped, or cancelled generations leave immediately so a waiting stream can take the slot on the next token boundary.
floor
residency
entitlement
independence
Pilot gate
One four-node pod validates or kills the thesis before scale capital.
$110K experiment
Four 512 GB nodes, fabric, rack, instrumentation, and the real serving path.
Throughput pass bar
Hold at least two reserved or five pooled streams above the 12 tok/s floor.
Commercial proof
Three to five design partners willing to pay per seat after the pilot.
Design partners
Bring CueCode to teams with real agentic coding load.
The first deployment should measure batching efficiency, duty cycle, time-to-first-token, tiering latency, and the buyer appetite for private per-seat inference.