Investor updates
Published as milestones land. Every entry is dated, mapped to a module in the register, and closes with what it means for investors.
Updates published
14
Modules reporting
7
Most recent
31 Jul 2026
Reporting cadence
As milestones land
Foundation model and inference engine
The pilot checkpoint passed all twelve blocking evaluations, including the multi-step tool-use suite that failed in April. Promotion to the 70B production pretraining run is approved.
Investor takeawayThe main technical risk on the foundation model is now retired; remaining risk is compute schedule, not capability.
Scheduler rework lifted sustained training utilisation from 68% to 91%, reducing cost per training token by roughly a third against the original plan.
Investor takeawayLower unit training cost extends the runway of committed capital without changing the milestone dates.
Data ingestion and synthetic corpus pipeline
Policy-constrained synthetic generation is running for scarce industrial domains, with quality scoring gating every batch before it reaches the curriculum sampler.
Investor takeawayReduces dependence on third-party licensed data and the pricing risk that comes with it.
Every document now carries a signed licence attestation and lineage hash that follows it into model checkpoints, covering the full 4.2PB curated corpus.
Investor takeawayEnterprise procurement and EU AI Act diligence on training data can be answered with evidence rather than assertion.
Multimodal reasoning and vision stack
Document answers now cite back to the source pixel region, validated across contracts, engineering drawings and scanned records.
Investor takeawayMakes the document workloads sellable into regulated buyers who require traceable answers.
Image and document encoders now share the foundation model token space, removing the separate serving path previously planned.
Investor takeawayOne fewer system to run and maintain; modest reduction to ORACLE's forecast operating cost.
Autonomous agent orchestration layer
Typed schemas and permission scopes are in place across internal and third-party tools, with every call recorded in the replay trace.
Investor takeawayBreadth of integrations is the practical moat for agent products; this is ahead of the plan for the period.
Workflows now survive restarts and partial failures without state loss, with exactly-once semantics on side-effecting steps.
Investor takeawayMoves HELIX from demo-grade autonomy to something an enterprise can put in a back-office process.
Safety, alignment and compliance framework
Per-tenant inference-time rules are enforced across the staging environment, with violations logged to the compliance evidence pipeline.
Investor takeawayCompliance becomes a shipped control rather than a documentation exercise at audit time.
The full control set is mapped and internally reviewed, ahead of external assurance work scheduled for 2027.
Investor takeawayShortens enterprise security review cycles, which is typically the longest step in converting a pilot.
Edge deployment and compute fabric
Two launch partners reviewed and approved the compute fabric architecture covering reserved, spot and on-premise capacity.
Investor takeawayDesign partner sign-off de-risks the largest remaining infrastructure build in the programme.
The distilled model reaches 88% of pilot-model task quality on constrained hardware, validating the edge deployment thesis.
Investor takeawayOpens the on-premise and edge segment without a separate model programme.
Enterprise interface and integration suite
The usage schema is locked, which unblocks the billing build and the design of usage-based commercial terms.
Investor takeawayRevenue mechanics are now defined ahead of the 2028 general availability milestone.
A cohort across logistics, utilities and financial services is validating the workspace and connector priorities.
Investor takeawayEarly demand signal for the commercial surface, and the source of the first connector roadmap.
Formal reports are published every quarter, with a downloadable PDF for each module and one for overall programme progress.
Programme-level
The pilot checkpoint passed all twelve blocking evaluations, including the multi-step tool-use suite that failed in April. Promotion to the 70B production run is approved for August.
38% of committed capital deployed against a 41% plan. Underspend is concentrated in VERTEX hardware validation, which shifts to H2 following a revised partner schedule.
A logistics group and a regional utility joined the ATLAS design partner programme, taking the cohort to five and covering three regulated verticals.
Every document entering the training corpus now carries a signed licence attestation and lineage hash traceable into model checkpoints.
The full control set has been mapped and internally reviewed, ahead of external assurance readiness work scheduled for Q1 2027.
Scheduler improvements lifted sustained utilisation from 68% to 91%, materially reducing the cost per training token ahead of the 70B run.
