Crusoe
Each of these is a real decision Crusoe is working through, written so you can make the call yourself — incomplete information, someone who pushes back, no answer key.
Crusoe designs, builds, and operates purpose-built, AI-optimized data centers, from gigawatt-scale campuses like Stargate in Abilene, Texas, to modular and distributed deployments, and runs hyperscale AI data center construction projects. Its development pipeline exceeds tens of GWs across multiple U.S. campuses. Products include Modular Data Center (Spark) and Power Distribution Center (PDC), and it operates multiple production sites manufacturing AI data center infrastructure. Its data centers are powered by on-site generation, stranded and flared gas, and grid and renewable PPAs, and it has power plant facilities across the USA. Crusoe Cloud targets developers, ML engineers, and technical builders. Strategic enterprise customers deploy AI/ML workloads on its high-performance GPU infrastructure, including NVIDIA A100, H200, GB200, B200 and AMD 350X / 355X. Customers run production LLM workloads on the Managed Inference platform, one of its fastest-growing product areas, without managing low-level infrastructure, and run containers on Crusoe Managed Kubernetes. Crusoe operates a global edge, backbone, and data center network for high-performance compute (HPC) clusters with GPUs.
Your first one is free. No card.
77 live right now.
The rest
Turn live endpoints into production habits before usage stalls
Pilot customers can deploy an endpoint yet still hold back production traffic and inference consumption. You must weigh a repeatable programme against each customer’s different reliability and value proof needs.
Commit to a dedicated endpoint while launch traffic is waiting
Launch-rehearsal users saw failed LLM responses, and the customer wants production traffic moved this week. You must weigh a fast dedicated cutover against the evidence and capacity commitments needed to make it safe.
Move the second workload before another manual capacity peak
The customer manually raises capacity before every weekly spike and is holding back its second workload. You must weigh faster adoption against the platform lead’s need to protect the live endpoint.
Trace a GPU charge to ranked causes without writing to the ledger
Billing engineers cannot quickly connect a disputed GPU charge to the workload events behind it. You must weigh faster investigations against evidence that an engineer can defend.
Forecast GPU demand without turning planning into a promise
Pilot customers cannot tell whether their next 30 days of GPU demand fits their current quota. You must weigh a useful planning signal against the risk of implying guaranteed capacity.
Fix twice-charged runtime in the ledger without mass credits
Some recovered Kubernetes workloads appear to have been charged for the same GPU runtime twice. You must weigh a fast customer correction against an audit trail Finance can defend.
Correct GPU charges as workloads move between projects
A share of GPU hours is landing on the project a workload left. You must weigh a fast correction against ledger integrity and the invoice deadline.
Route network alerts before stalled training burns more GPU time
Training jobs are stalling while responders assemble network evidence from separate views. You must weigh faster routing against evidence an engineer can trust before setting severity.
Prove host state before sensitive training begins
Security reviews are delaying sensitive training workloads despite available GPU capacity. You must weigh customer-verifiable evidence against protection of the fleet’s hardening details.
Choose a kernel rollback before reserved GPUs miss training
Some GPU hosts cannot rejoin Kubernetes after kernel hardening, blocking reserved training capacity. You must weigh a verified security baseline against an irreversible capacity decision under a customer deadline.
Stop rotation of healthy GPUs after brief telemetry gaps
Healthy GPU workers are being rotated after telemetry has already recovered. You have to protect fast failure response without draining usable inference capacity.
Surface repeat defects without weakening quality decisions
Recurring enclosure fit-up NCRs are taking days to connect to a defensible physical cause. You must weigh faster evidence correlation against the review needed before a build change is approved.
Design cooling connections that crews can replace without specialists
Dense-rack operators need cooling connections that can be verified and replaced without extending field work. You have to weigh hydronic performance, factory yield, and service access in one pilot-ready design.
Decide if a Spark L can ship without matching control records
A control cabinet in the module does not match the genealogy in its turnover record. You must weigh the shipment slot against evidence that the receiving operator can trust.
Fix the flow mapping so every loop matches within 2%
One cooling loop reports 18% less flow on the operator view than at the calibrated instrument. You must weigh a fast signal correction against the evidence needed to release a mission-critical module.
Route low-flow alarms while GPU workloads keep changing
Low-flow alarms reach the ROC without enough evidence to judge their thermal impact. You must weigh faster routing against evidence an engineer can safely trust.
Prove crews can isolate a dense module without specialist help
Operators cannot find one agreed isolation boundary in the turnover material for a high-density module. You must weigh self-service recovery speed against guidance that never outruns verified site controls.
Decide on cooling protection while turnover evidence conflicts
Low-flow alarms appeared during a GPU-load test even though the CDU reported recovery. You have to weigh the turnover date against evidence that the cooling protection will act safely.
Secure switchboard access while the module release stays on schedule
A technician can follow the lockout steps and still face an energized adjacent compartment. You have to weigh a fast factory release against a service boundary people can trust.
Route RDMA alerts while GPU jobs wait for a decision
Storage-backed training jobs are stalling while responders reconstruct context across several systems. You have to weigh faster routing against evidence an engineer can trust before acting.
Contain a leaked key in object storage without stalling training runs
A leaked object-storage key has no bounded containment path once it reaches a shared training environment. You must weigh rapid isolation against preserving active dataset and checkpoint access.
Hold tenant admission until failover tests show no shared RDMA path
A failover test exposed a possible cross-tenant RDMA path before 2,000 GPUs receive scheduled workloads. You must weigh cluster availability against evidence strong enough to commit an irreversible network change.
Stabilize file reads as PFC pauses stall GPU workers
Checkpoint reads are freezing for seconds on one Kubernetes cluster while PFC pauses build on its leaf pair. You must weigh a fast leaf-policy correction against lossless RDMA behavior for nearby storage traffic.
Compare candidate models without changing the live endpoint
Enterprise ML teams lack a safe way to judge a candidate model on production-like requests. You have to weigh useful live evidence against endpoint stability and GPU cost.
Commit scarce GPUs while a customer endpoint misses its timeout
A dedicated endpoint is missing its 1.8-second latency commitment before a customer traffic event. You must weigh a faster recovery against the throughput already promised to another deployment.
Separate long prompts before short requests stall
Short questions are queuing behind document-sized prompts, pushing first-token latency past the customer’s interactive target. You have to weigh responsive routing against GPU efficiency and a safe rollback path.
Match billing to approved terms while Finance keeps final say
Exceptions between approved GPU terms and billing records are staying open after buyers sign. You must weigh faster review against keeping every financial judgment with Finance.
Settle how reserved GPUs are priced so quotes reconcile to Stripe
Enterprise buyers receive GPU comparisons that change when they ask about utilization or workload flexibility. You have to weigh a simpler buying decision against assumptions Finance can defend after the commitment is signed.
Close the prepay without weakening the guarantee on unused capacity
An $18 million prepay is at risk because the buyer wants workload flexibility that changes the commitment behind it. You must weigh cash certainty and deal speed against margin protection, revenue treatment, and a precedent future buyers will cite.
Fix approvals for prepaid quotes without loosening margin controls
Standard prepaid GPU quotes are sitting for days despite matching approved commercial guidance. You have to weigh a faster approval path against controls over margin, billing, and exceptions.
Reconcile site status before cluster dates split again
The site-expansion record conflicts across Jira and Confluence while enterprise ML engineers await a first-cluster date. You must weigh faster status visibility against keeping readiness, dates, and priority with Program Managers.
Settle on one date for engineers while teams keep decision rights
Enterprise ML engineers are receiving conflicting dates for a new GPU cluster because readiness work is split across teams. You must weigh a simple customer path against evidence each infrastructure owner needs before committing.
Choose the GB300 handoff while commissioning evidence conflicts
A GB300 cluster is allocated to a customer, but repeat commissioning runs disagree on whether it is ready. You must weigh the delivery commitment against evidence that may only reveal its meaning after cutover.
Align commissioning gates before a ready cluster misses its handoff
The customer cluster is waiting because its readiness record disagrees with the daily review. You must weigh fast ownership clarity against adding process that engineers will not maintain.
Structure discovery records before proposal assumptions harden
Technical discovery details are reaching proposal work with open questions about Kubernetes and committed GPU capacity. You must weigh a faster usable next step against source-linked seller review of every commitment.
Pick which SIs to back without blurring who owns each account
Partners are reaching enterprise AI teams but cannot yet carry a consistent capacity and Kubernetes story. You must weigh partner reach against scarce technical support and clear account ownership.
Commit capacity for the revised workload despite unproven assumptions
A $120M compute agreement is at risk because the buyer’s revised Kubernetes architecture no longer matches the capacity originally planned. You must weigh customer certainty against commitments that Data Center Infrastructure and Energy can stand behind.
Reshape GPU terms while procurement stalls the Kubernetes contract
A $12M annual commitment is stalled because training capacity and inference burst use are bundled together. You must weigh a clear commercial commitment against the flexibility ML engineers need.
Correlate InfiniBand failures so qualified racks return sooner
GPU racks stay under review when end-to-end InfiniBand failures cannot be tied to their build history. You must weigh faster capacity return against evidence an engineer can trust for a physical-cause decision.
Qualify 20 hosts across four racks before they join the fleet
New GPU hosts can pass local checks without evidence that PCIe, InfiniBand, and NVMe hold up together across a rack. You have to weigh faster fleet admission against a qualification method manufacturing can repeat and operations can trust.
Decide whether to hold the cluster while PCIe resets stay unexplained
Workers disappear from long-running multi-node jobs even though ordinary burn-in runs pass. You must weigh committed capacity against a failure mode that becomes costly after installation.
Fix the link training on serviced hosts without idling healthy capacity
Recently serviced GPU hosts are missing InfiniBand qualification and cannot rejoin distributed workloads. You must weigh fast capacity recovery against evidence that the physical correction will hold.
Route conflicts on cooling to owners instead of debating them in review
Cooling-interface changes are appearing differently across the schedule, change log, and Design-to-BoM record. You must decide what the system can reconcile and what program managers must still judge.
Prove high-density readiness before the module reaches operations
Infrastructure managers cannot tell whether factory evidence covers the high-density load they plan to run. You must weigh a fast turnover against evidence operations can rely on.
Set the shipment configuration before customs filing closes
Conflicting module records could delay a Spark shipment and its commissioning start within 21 days. You must weigh the committed site sequence against a configuration record that can withstand export review.
Reconcile cooling records before the work order opens
A Spark cooling package has conflicting released records, putting the next kitting run at risk. You must weigh release speed against the cost of building the wrong configuration.
Rank inference accounts for review while CSMs still judge every outreach
Inference accounts are reaching escalations before their full technical history is visible in a review. You must weigh earlier risk signals against evidence a CSM can verify and act on.
Shape readiness reviews before reserved GPUs sit unused
Ten pilot customers are approaching reserved GPU windows without a shared definition of production readiness. You have to weigh a repeatable customer programme against the technical variation each Kubernetes workload brings.
Commit the GPU cutover while latency evidence stays disputed
Long-running Kubernetes tests still produce intermittent latency spikes as the customer’s current capacity window closes. You must weigh launch capacity against the cost of committing before the customer can trust sustained performance.
Restore scaling before the next inference traffic burst
Requests queue after traffic bursts while healthy GPU capacity remains unused. You must weigh a fast configuration change against a production setup the customer can confidently run.
Rank serving setups so pilots stop burning days on dead ends
Similar inference workloads repeatedly converge only after several failed configuration trials. You have to weigh a useful recommendation surface against sparse telemetry and the risk of overconfident predictions.
Decide on SGLang cutover while streamed tokens arrive out of order
Thirty-seven token-order violations appeared in 2.4 million sustained-load sessions on the committed host group. You have to weigh a contractual launch window against evidence that is expensive to verify and hard to reverse.
Tune vLLM routing before the nightly batch misses its window
Long-context requests push P95 time to first token to 4.6 seconds during the nightly batch. You have to weigh a reversible routing change against the risk of disrupting a live inference window.
Cluster storage feedback before recurring training friction disappears
Storage feedback from Kubernetes and Slurm may describe the same training problem without a shared evidence trail. You must decide what AI can group, what PMs must verify, and how the team will run it afterward.
Order GPU jobs without changing reserved capacity
Pilot customers cannot tell which queued training job should receive scarce GPUs first. You must weigh clearer workload priority against consistent behavior across Kubernetes, Slurm, and reservation commitments.
Commit the cluster while checkpoint recovery remains disputed
Checkpoint recovery stalls appear only under representative distributed workloads, days before a fixed customer handoff. You must weigh the customer’s training window against evidence that is expensive to reproduce and hard to dismiss.
Clarify node status so training jobs stop being canceled
ML engineers are canceling jobs when a temporary node rotation looks like lost capacity. You have to weigh a clearer customer signal against the risk of prompting action during active health work.
Decide what the AI may run in triage until an engineer approves
Repeated integration-test failures are moving GPU-host release dates for ML engineers. You must decide which evidence gathering can run automatically and where an engineer must take over.
Build a gated path for sharing GPUs without touching reserved capacity
Enterprise ML engineers cannot trial shared scheduling on a validated cluster, even as managed alternatives become more visible. You must weigh a useful first deployment against the validation burden and capacity protections around it.
Decide the GPU cutover while isolation evidence conflicts
The new host group passes short tests but shows tenant interference after sustained multi-node load. You must weigh committed capacity against evidence that is costly to reproduce and verify.
Fix the Ansible rendering of network settings before training starts
Newly promoted hosts lose network throughput only when multi-node training traffic begins. You must weigh a fast configuration rollback against confidence that the deployment path remains safe to promote.
Decide what AI may assemble while engineers confirm every cause
GPU dashboards can go stale even while Kubernetes workloads continue running. You must decide what AI can assemble automatically and what engineers must verify before acting.
Expose log costs by workload without heavier onboarding
Pilot customers cannot connect a managed logs increase to the Kubernetes workload that caused it. You have to weigh useful attribution against a low-touch, tenant-safe telemetry path.
Choose telemetry limits while shared GPU queries slow
Unrelated Kubernetes customers are waiting on queries while one training cluster floods the shared telemetry path. You have to weigh fast recovery for everyone against losing the labels needed to investigate scarce GPU capacity.
Restore missing logs while Kubernetes onboarding still looks healthy
Affected Kubernetes customers can install the agent successfully and still get no searchable logs. You must weigh a fast onboarding correction against preserving tenant attribution and access controls.
Price the lease-back before operations inherit its costs
The first powered-shell deployment has no shared view of factory cost, recurring service cost, and return before Q1 2027. You have to weigh a usable first forecast against assumptions that may change as commissioning closes.
Decide the draw before turnover locks the module cost
Remaining cooling, customs, and commissioning costs do not agree across the project records. You must weigh a timely turnover against a draw that Accounting can defend.
Catch overruns in module costs before the month-end close
Spark modules are landing on schedule while material costs surface above forecast only at close. You have to weigh timely margin visibility against the control needed for inventory valuation.
Reconcile resiliency status before the weekly review loses trust
Three records disagree about whether a storage dependency is blocked or complete. You must decide what AI can reconcile and what the Program Manager must verify.
Settle who owns failures once a transfer has started
Enterprise ML engineers lack a clear path for moving object data across clouds and handling failures. You must weigh a useful first launch against readiness that support and infrastructure can sustain.
Commit file capacity before the provider allocation expires
A provider allocation closes in 10 business days while shared file system demand is rising faster than the approved plan. You must weigh protected training capacity against an irreversible commitment that could leave storage idle.
Clear blockers in Jira before the Storage deployment review
Storage deployment reviews keep uncovering dependencies with no owner or current date. You have to weigh a faster readiness rhythm against the burden of keeping Jira accurate.
Trace object stalls while GPU jobs wait
Object reads are slowing active ML jobs even when the data remains reachable. You must decide what AI can assemble automatically and what engineers must verify before acting.
Archive checkpoints without leaving ML engineers guessing at recovery
Completed checkpoints keep consuming shared file capacity because ML engineers do not trust an archived copy until a restart proves it. You have to weigh faster capacity release against the recovery evidence a training owner needs.
Contain object listings without freezing customer data migrations
Some ML engineers can see object prefixes outside their tenant after credential rotation. You must weigh immediate tenant containment against keeping authorized dataset movement available.
Stop GPU pauses in NFS reads without reverting the NVMe/TCP change
NFS checkpoint reads are pushing GPU workers into visible pauses during a training run. You must weigh a reversible path change against the throughput gain it was meant to preserve.