Maryanne Baines has spent the last few years in the trenches of GPU-as-a-service—scoping clusters, negotiating power and land, and watching customers pivot from desperate capacity grabs to deliberate platform choices. In this conversation, she unpacks why the “rent-a-GPU” play is fragile, how operators can move up the stack without antagonizing hyperscalers, and what the real levers are for margins, moats, and financing. Expect grounded stories tied to the market’s most telling numbers—like 77% revenue concentration with two customers and a quarter featuring $1.21B in revenue, a $290.5M net loss, and $2.9B in capex—and a roadmap for where neoclouds must evolve next.
Big picture, how do you see “neocloud” GPU-as-a-service evolving over the next 2–3 years, and what moments from your own deals or deployments show where the market is headed? Please share concrete metrics, customer behaviors, and a step-by-step example of a recent shift.
The arc bends from pure capacity brokerage to opinionated platforms. Early on, customers took whatever GPUs they could get; that desperation funded builds like quarters with $2.9B in capex because demand outstripped supply. Now I’m seeing buyers ask for curated training toolchains, inference SLAs, and domain stacks. One recent shift: a team that had been chasing spot capacity moved to a multi-region reservation with a unified runtime. Step-by-step, we aligned on three moves: codify a standard image and data ingress pattern, attach a managed inference gateway, and set clear job start targets. The behavioral tell was willingness to trade some price flexibility for predictability—mirroring the market where a company can post $1.21B in quarterly revenue while still running a $290.5M net loss because it’s racing to platform scale.
McKinsey calls the model “inherently commoditized.” What specific levers have you seen actually move margins—pricing, SLAs, networking, data locality, or software? Walk me through a case where differentiation worked, including numbers, timelines, and the before-and-after impact.
SLAs and software layers moved the needle most. We took a customer off undifferentiated rentals and wrapped their workloads with a managed orchestration layer and a firm job-start window. In 60 days, queue anxiety evaporated, and upsells followed because they trusted the pipeline. Pricing alone was a race to the bottom; adding a domain stack lifted attach rates. The dynamic resembles why investors expect neoclouds like CoreWeave, Nebius, and Crusoe to move up-stack: pure hardware resale, often Nvidia, pushes you into brutal price comps unless you anchor on reliability and workflow.
Many neoclouds resell access to Nvidia gear. Where can operators create moat-like value beyond hardware—tooling, data pipelines, orchestration, or model optimization? Give a practical story with metrics and a step-by-step breakdown of how those layers changed win rates.
The moat forms where data meets orchestration. We won a bake-off by bundling an ingestion pipeline, a reproducible build system, and a runbook for drift. Step-by-step: standardized data landing zones, enforced lineage in the training loop, then exposed a single control plane for both training and inference. That bundle collapsed onboarding time and reduced failed jobs. The proof is market-wide: when customers see operators evolving into training tools, inferencing services, and domain stacks, they stop shopping by GPU model and start buying outcomes.
Moving up the stack risks competing with hyperscalers who are also top customers. How do you structure partnerships, product roadmaps, and account plans to avoid channel conflict? Share an anecdote, concrete thresholds you use, and the decision tree you follow when overlap appears.
We draw a hard line: horizontal primitives are partner-first; vertical stacks are ours where hyperscalers are less effective or less welcome. The decision tree is simple: if a feature looks generic, we co-sell; if it’s domain-specific and compliance-heavy, we lead. An anecdote: we declined to productize a generic feature after discovering overlap with a hyperscaler that, for another operator, represented 62% of revenue in a year—too risky to antagonize. Our threshold is: if a top partner crosses half of a segment’s revenue, we default to complement, not compete.
CoreWeave had 77% of 2024 revenue from two customers, with Microsoft at 62%. What concentration guardrails make sense in this market? Describe your target mix, covenant thresholds, and a real example where you rebalanced exposure without losing growth.
I set soft guardrails at no more than half of revenue from one logo and aim to cap the top two at under three-quarters—because 77% from two creates fragility. In practice, we hedge with staggered terms and reservation mix. We once faced a spike from a single anchor; we rebalanced by prioritizing mid-market reservations and insisting on take-or-pay in a few smaller deals. It slowed one quarter’s top-line but reduced the chance of cliff risk if a hyperscaler rebases spend.
In Q2, CoreWeave reported $1.21B revenue (+207% YoY), a $290.5M net loss, and $2.9B in capex. How do you justify that capex curve? Walk through the model you’d use—utilization assumptions, payback periods, WACC, and a step-by-step sensitivity that would change your plan.
The thesis is scale-before-stability: front-load capex to capture share during constrained supply. My model ties new racks to contracted demand with buffers for burst. Step-by-step: anchor reservations, then layer opportunistic capacity; monitor utilization and queue times; if churn or start latency worsens, pause expansions. When a quarter shows $2.9B in capex with a $290.5M loss against $1.21B revenue, it’s a bet that platform layers will monetize that footprint. If price pressure accelerates or utilization dips, I pivot to deferring builds and doubling down on software margin.
CoreWeave raised $25B in debt and equity in 18 months. What financing stack best fits GPU buildouts today—term loans, ABS on contracts, leases, or equity? Share concrete costs of capital, covenant examples, and a timeline of how you’d stage funding against bookings.
I like a ladder: equity to seed the first tranche, then contract-backed facilities as bookings firm, and leases to smooth working capital. Raising $25B in 18 months tells you the stack must be diversified. Covenants should align with utilization and reservation coverage; tie drawdowns to booked capacity. The timeline is bookings-led: commit equity to unlock supply, add ABS-like structures once take-or-pay is inked, and keep dry powder for power and land that can’t wait.
Microsoft is turning away some training jobs in favor of higher-margin inference. How do you help customers choose between training and inference allocations? Provide a real cost-per-token, latency, and throughput example, plus the steps to optimize mix across regions.
The trade is margin versus innovation tempo. When a hyperscaler prioritizes inference, it signals where the economics sit today. We guide customers by mapping their cadence of model updates to inference SLAs and deciding what can shift to regions with headroom. Steps: profile demand, classify jobs by urgency, place inference near users for latency, and batch training where queues are shorter. The result is fewer blocked launches and a happier finance team.
CoreWeave’s megalease with Applied Digital and the $9B Core Scientific deal point to a power-first strategy. How do you secure power and land at speed? Walk me through site selection, interconnect lead times, power pricing, and a checklist you use to de-risk builds.
Power is king; land is a close second. Deals like a megalease or a $9B acquisition are really grid strategies. We triage sites by available megawatts, interconnect timelines, and fiber diversity; then we lock pricing bands to hedge volatility. My checklist: confirm substation upgrades, validate water or air-cooling viability, pre-negotiate interconnect, and secure multi-path fiber. If any leg wobbles, the schedule slips and your platform roadmap follows.
If GPU supply catches up, some neoclouds may fade. What durable demand signals do you track—model size trends, context windows, multi-tenant inference loads? Give a recent data-backed example and a step-by-step scenario plan if prices fall 30% in 12 months.
I watch sustained inference concurrency and the shift to domain stacks—because that’s sticky. The market already hints at this with operators steering toward inferencing services and industry-specific platforms. If prices drop 30%, step-by-step: lock in longer-term reservations, push software-led bundles, and prune unprofitable SKUs. The goal is to protect margin via value, not just price.
On niches where hyperscalers struggle, which sectors—financial services, healthcare, public sector—offer the best fit? Share a concrete workload, compliance posture, and an anecdote showing how specialized services beat generic cloud, including the specific KPIs that mattered.
Financial services loves low-latency inference with rigorous audit trails. We won by pairing an inference gateway with a domain stack and pre-approved controls. The KPIs weren’t just latency—they included explainability workflows and repeatable deployment patterns. This is why McKinsey points to niches and domain stacks as defensible paths; you meet customers where hyperscalers are less welcome.
Serving startups as a launchpad can be risky. How do you pick winners and structure contracts—prepaid credits, usage floors, or warrants? Provide a real funnel example, cohort metrics over 12–24 months, and the steps that turned a small logo into enterprise scale.
I look for early signs of product-market fit and a cadence of launches. We pair prepaid credits with usage floors to align incentives. The step-up path: get their first model live, lock a reservation when growth appears, then layer managed services. This mirrors the “launchpad to AI-native enterprise” pathway—start small, grow into platform spend without overexposing concentration.
If consolidation is coming, what makes a neocloud worth buying—power contracts, data centers, network, software IP, or customer base? Tell a story of a deal you would do or pass on, with valuation ranges, synergy math, and an integration playbook.
I’d buy power and network first, software IP second, and customers last if they’re concentrated like 77% from two logos. I passed on one with thin interconnect and no domain stack; the synergy wasn’t there. The playbook I favor: keep the best-of-breed runtime, migrate customers gradually, and standardize observability. Consolidation works when you inherit durable moats, not when you just add more racks.
What KPIs best show real differentiation—SLA adherence, job start latency, queue times, cost per trained token, cost per million inferences? Share live benchmark numbers you’ve seen, how you collect them, and a step-by-step method to improve any one KPI by 20%.
SLA adherence, job start latency, and queue times are the heartbeat. We collect them from the control plane and surface them in customer dashboards. To improve job start latency by 20%, step-by-step: pre-warm pools, optimize scheduler hints, pin hot datasets, and reserve interconnect paths. Those moves convert platform reliability into pricing power.
For moving up the stack, which software layers would you build first—training tooling, inference gateways, vector stores, or vertical stacks? Describe a phased roadmap with milestones, hiring plans, and hard metrics you’d hit before funding the next layer.
I’d build inference gateways first, then training tooling, then vertical stacks. Milestones: production-grade gateway usage and SLA wins; reproducible training runs; finally, a domain stack that unlocks co-selling. Hiring follows that sequence: SREs and networking first, MLOps engineers next, domain experts last. Funding gates hinge on attach rates and renewal—because that’s the path from commodity to platform.
Do you have any advice for our readers?
Treat hardware as table stakes and design your moat in software, power, and trust. Watch your concentration—if one logo nears half your revenue, adjust. Move up the stack where hyperscalers are less effective or less welcome, and partner everywhere else. Most of all, build a roadmap that can survive price compression by delivering outcomes customers can’t get from a bare rental.
