Goliath Super Intelligence
Goliath SI / Infrastructure
COMPUTE ESTATEInfrastructure

One hall underground, 3,057 PB, doubling every 49 days

284 NVIDIA Vera Rubin and 470 NVIDIA B300 Blackwell Ultra accelerators in service at a single underground location, behind 3,057 PB of usable tiered storage. Storage doubles every 49 days. Compute multiplies 1.4 times every 26 days. The deep ice hall is commissioning now and subsea follows it.

3,057 PBUsable, all tiers
49 daysStorage doubling cadence
26 daysCompute 1.4x cadence
754Accelerators in service
The site

Underground, one location, and quiet about it

Goliath SI computes from a single location and that location is below grade. Rock is a better neighbour than a business park: it takes the thermal load, it takes the acoustic load, it holds a stable ambient year round, and it removes the entire class of risk that comes from being visible from a road. The hall was built into existing excavated space rather than poured on a greenfield site, which is also why it came up in months rather than years.

We operate under the radar deliberately. There is no plaque, no campus, no tour, and the site is not named here. What is published is what the system does and what it is made of, because that is what a customer running an autonomous fleet against their own estate needs to evaluate. Where it sits is not part of that evaluation and it is not going to be.

The fourth storage tier is the one exception to the single location rule, because disaster recovery that shares a mountain with the thing it is recovering is not disaster recovery. Nothing computes there. It holds immutable copies and it is reachable by nothing in the hall that could delete from it.

The public edge is also not ours: goliathsi.com is served from Cloudflare, which keeps the marketing surface and the compute estate on separate infrastructure with no shared credentials.

Accelerators in service
284 x NVIDIA Vera Rubin VR200 (2 x NVL144 racks, 288 installed, 4 cold spares)470 x NVIDIA B300 Blackwell Ultra (6 x GB300 NVL72 racks plus one HGX B300 island, 472 installed)48 x NVIDIA H200 SXM (4 x HGX nodes, evaluation and regression)16 x NVIDIA L40S (visualisation, encode, light inference)8 x NVIDIA RTX 6000 Ada (operator workstations, adapter synthesis)
Host and control plane
NVIDIA Grace CPU (Vera Rubin and GB300 superchip hosts)AMD EPYC 9005 Turin (storage, control plane, ingest)Intel Xeon 6 (management, licence anchored workloads)NVIDIA BlueField-4 CMX (context memory, tier 1 storage path)NVIDIA BlueField-3 DPU (tenant isolation, storage offload)2N UPS on the control plane, N+1 on computeOut of band management on a physically separate switch plane
Fabric
NVLink switch domain inside every rackNVIDIA Quantum-X800 InfiniBand XDR spineNVIDIA Quantum-2 NDR leaf (H200 and storage islands)Rail optimised topology, eight rails per nodeNVIDIA Spectrum-X Ethernet for north-south and object traffic800G OSFP uplinks, LPO optics inside the rowGPUDirect RDMA end to end, no host bounce on the training pathSHARP in-network reduction on the collective path
Power and cooling
11.4 MW contracted at the surface feed, 7.8 MW drawn at steady stateDirect to chip liquid cooling, 6 CDUs per row, N+1132 kW per NVL144 rack, 121 kW per NVL72 rackRock loop as the primary heat sink, no evaporative waterRear door heat exchangers on the air cooled islandPUE 1.09 measured across the last full monthDual utility feed plus on site generation for the control plane
Software and scheduling
Slurm for training, Kubernetes for inference and servicesenroot and pyxis for container execution under SlurmNVIDIA NeMo and Megatron-LM for pretraining and fine tuningTensorRT-LLM and Triton Inference Server for servingNCCL 2.23 with topology aware rail assignmentDCGM and NVML telemetry into first party analyticsCustody warden issuing short lived credentials, no static secrets on a node
Storage

Four tiers, 3,057 PB usable, sized by what the data is for

Storage is tiered by purpose rather than by how often it is touched. A KV cache, a training corpus, an archived video pull and an immutable backup have nothing in common except that they are all bytes, and putting them on the same medium makes all four worse.

The top of the stack is memory, not a filesystem. Inference context, the LLM KV cache and the checkpoint a run is actively writing sit on NVIDIA STX with BlueField-4 CMX, GPU adjacent and addressed as CXL class memory rather than as a mount point. That is what lets a long context request reuse its cache instead of recomputing it, and it is the single change that most affects what a Sling costs to run.

Beneath that, VAST Data carries the working set on all flash NVMe and is sized in write bandwidth before capacity, because when the fleet checkpoints together the filesystem has to absorb the write inside the window before the next step. Bulk capacity sits on MinIO AIStor ExaPOD over enterprise HDD, addressed as S3 in the same namespace, and the fourth tier is immutable and geographically separate.

Every tier is checksummed on ingest and on read, and the provenance chain is written append only, so the artifact behind a published claim is still there, and still verifiable, when somebody asks in three years.

Tier 1 — GPU context and ultra fast storage
NVIDIA STX with BlueField-4 CMX. NVMe and CXL class memory, 6 PB usable, GPU adjacent. LLM KV cache, inference context, hot datasets and the model checkpoints a run is writing right now. Addressed as memory rather than as a filesystem, so a cache hit never leaves the accelerator complex.
Tier 2 — high performance shared storage
VAST Data on all flash NVMe, 66 PB usable, 3.3 TB/s aggregate write measured against a full fleet checkpoint. Training datasets, active AI workloads, the model repository and anything needing parallel file access across the fleet. This tier exists to make the collective stop waiting.
Tier 3 — main capacity storage
MinIO AIStor ExaPOD on enterprise HDD, 1,185 PB usable, S3 addressed and erasure coded. Bulk datasets, historical AI training data, video, images, documents, embeddings and archives. Promotion to tier 2 is a policy change inside one namespace rather than a migration.
Tier 4 — disaster recovery
1,800 PB geographically separate, immutable, write once with object lock and an offline copy. Protected datasets, the custody vault, the AEGIS adjudication record and long term retention. Nothing computes at that location and no credential in the hall can delete from it.
Metadata and ledger
NVMe backed key value store for the index, the tether state and the credential leases, replicated three ways inside the hall, with the ledger written append only and sealed hourly into tier 4.
Storage architecture
NVIDIA STX with BlueField-4 CMX, GPU context tierCXL class memory addressing for KV cache and hot setsVAST Data all flash NVMe, parallel file and objectNVMe over fabrics on the training path, GPUDirect Storage end to endMinIO AIStor ExaPOD, S3 addressed enterprise HDD capacityErasure coding on tiers 3 and 4, no RAID rebuild windowsObject lock and write once retention on the DR tierChecksum on ingest and on every read, all tiersOne namespace across tiers 2 and 3, policy driven promotionAppend only provenance ledger, sealed hourly off site
Cadence

Storage doubles every 49 days, compute multiplies 1.4x every 26 days

Both cadences are build rates, not forecasts of demand. Storage has doubled on a 49 day interval through the current estate, which at 3,057 PB usable means the team lands 43.6 PB of usable capacity a day, every day, including the days when nothing arrives and two days of capacity go in at once. Compute multiplies 1.4 times every 26 days, which at 754 accelerators in service is just under ten accelerators a day.

The storage cadence survives because nothing about it is bespoke. Capacity arrives as pre-burned-in enclosures with the controller headroom already paid for, joins an existing namespace instead of forming a new one, and is erasure coded rather than rebuilt, so there is no rebuild window to schedule around. Tier 3 absorbs most of each doubling, tier 2 is sized against the collective write rather than the dataset, and tier 1 scales with the accelerator count because it rides the DPU. A doubling is therefore a receiving dock problem and a cable problem. It is not a filesystem problem, and that is the only reason it can be done on a timer.

The compute cadence is harder and slower on purpose. 1.4 times every 26 days is the fastest rate at which a new rack can be accepted, burned in, rail mapped, folded into the scheduler and proven against the regression suite without the fleet getting less correct while it gets larger. We have run it faster. The evaluation record got worse, so the rate came back down and has stayed there.

Carried forward, the storage cadence reaches about 40,700 PB by the end of May 2027 and about 534,000 PB inside a year. The compute cadence reaches about 8,050 accelerators over the same six months. Those are what the cadences imply rather than what is installed; what is installed is the figure at the top of this page, and it is the only one we will ever quote as current.

Power follows at roughly 1.35 times every 26 days, which the surface feed cannot carry past the second quarter of 2027. That is not a problem to be solved at this hall. It is the reason the next two halls are not at this hall.

Storage, as the 49 day cadence implies
now · 3,057 PB usable2027-01-04 · 6,114 PB2027-02-22 · 12,228 PB2027-04-12 · 24,456 PB2027-05-31 · 48,912 PB2027-07-19 · 97,824 PB2027-09-06 · 195,648 PB2027-10-25 · 391,296 PB2027-12-13 · 782,592 PB
Accelerators, as the 26 day cadence implies
now · 754 in service2026-12-12 · 1,0562027-01-07 · 1,4782027-02-02 · 2,0692027-02-28 · 2,8972027-03-26 · 4,0552027-04-21 · 5,6772027-05-17 · 7,9482027-06-12 · 11,1272027-07-08 · 15,5782027-08-03 · 21,810
What makes a doubling routine
Enclosures arrive burned in, no on site qualificationController and switch headroom bought one doubling aheadCapacity joins an existing namespace, never a new oneErasure coded expansion, no rebuild windowTier 3 absorbs the bulk of every doublingTier 1 scales with the accelerator count, not separatelyReceiving dock and cable plant are the rate limit, by designTwo days of capacity can land in one day without a change window
Deep ice and subsea

The next two halls are colder and harder to reach

The deep ice hall is commissioning now. Ice is the cheapest heat sink on the planet and it does not need water, a cooling tower or a chiller plant: a closed loop into a stable sub-zero mass carries the thermal load that a surface site would spend a third of its power moving. The hall is designed to run with no evaporative water at all and with a PUE that the surface feed cannot reach in any season. First racks are on site and under acceptance. It is not carrying production work yet, and this page will say so when it is.

Subsea is next and it is close. A sealed pressure vessel on the seabed has the same advantages and two more: the ocean is an effectively unlimited thermal reservoir, and a vessel that nobody can reach without a support ship is physically secure in a way that no fence achieves. The trade is serviceability, which is why the subsea programme is built around pods that are not serviced at all during their deployment, with capacity added by placing another pod rather than by opening one.

Across all three, the storage architecture is identical and the namespace is one namespace. A pod in the water, a rack in the ice and a rack in the rock present the same tiers to the same scheduler, which is what makes the 49 day cadence a logistics question rather than an architecture question. The only thing that changes between halls is where the heat goes.

Hall programme
Underground hall · in production · 754 accelerators · 3,057 PBDeep ice hall · commissioning · first racks on site under acceptanceSubsea pods · build · unserviced deployment, capacity added by placementDR tier · geographically separate · immutable, offline second copyOne namespace across every hall, same four tiersNo evaporative water at any hall
Precedent

The cadence is published because it becomes the reference

We publish the build rate for a reason that has nothing to do with marketing. A smaller AI company planning its own estate has almost nothing reliable to steer by: vendor roadmaps slip, benchmark pages describe clusters nobody can buy, and announcements describe intent rather than installation. A real cadence from a real estate, with the figure that is installed separated from the figure the cadence implies, is a compass. Ours is the one that gets used that way, and we would rather it be accurate than flattering.

That is also why the distinction is enforced on this page. 3,057 PB is installed and usable. 782,592 PB is arithmetic. Anybody planning against our numbers should be able to tell which is which at a glance, because we have watched what happens to companies that could not.

The question

How do you tell your storage team to double down on petabytes of data each month?

That is what we asked. Three years ago.

The answer was not a product and it was not a vendor. It was taking every bespoke decision out of the path: one namespace, pre-burned-in enclosures, headroom bought a doubling ahead, erasure coding instead of rebuilds, and a receiving dock sized for the day two deliveries arrive at once. Once none of those decisions had to be made again, the cadence stopped being a project and became a timer.

Monthly is what we asked for. 49 days is what came back, and 49 days is what has held.

Related