Neoclouds & Cloud Computing

Routing requires understanding query difficulty, model performance across domains, and rapid onboarding of weekly model releases; not all neoclouds (Fireworks, Together, Nebius) can build true routing; war unwon, technical moat matters.

Former Bitcoin mining firms with locked-in power access are repurposing infrastructure for AI compute, creating a new 'neocloud' sector that captures the spread between crypto mining economics and AI data center margins.

A new layer of capital-efficient inference clouds is emerging, monetizing open-source models via router/RL fine-tuning products. They grow nearly as fast as frontier labs but burn little cash (Rule of 40 extremes), creating a durable, high-margin infrastructure layer that expands total addressable compute demand.

Microsoft Azure's 43% YoY cloud revenue growth demonstrates that AI model training and inference workloads are now the core growth engine for hyperscalers; similar dynamics at Google Cloud and AWS suggest a multi-year capex supercycle for cloud infrastructure.

Google Cloud's 82% growth and negative FCF reflect market fear about $200B+ capex payback period, yet renting compute to frontier labs has been highly profitable; the real variable is whether token-maxing enterprises sustain demand or CIO budget clampdowns create air pockets.

Frontier labs (Google, Anthropic) are paying 2x spot prices ($900M/month for 110K GPUs) to neocloud providers like SpaceX for guaranteed capacity, efficiency, and security — signaling a structural shift where premium compute access is allocated via long-term contracts rather than spot markets, creating a new infrastructure layer with pricing power.

Cloud revenue growth is the key metric for hyperscalers; Google Cloud outpacing Azure and AWS re-accelerating, with positive estimate revisions expected as AI workloads drive cloud growth, though token monetization from frontier labs not yet visible in earnings.

Providers like CoreWeave, Nebius, Together AI, and Fireworks are positioning to license and serve open-weight frontier models like Kimi K3 on GB300 clusters, creating a new business model for GPU clouds beyond raw compute rental into model hosting and inference services.

XAI and OpenAI are building gigawatt-scale AI infrastructure from scratch within 12-18 months, matching or exceeding incumbent hyperscaler deployment velocity and reshaping the competitive landscape for AI compute supply.

Google Cloud's complexity (thousands of services) is now an asset because AI agents (MCP) can navigate APIs, provision resources, and manage operations programmatically; this turns breadth into a competitive advantage as enterprises adopt AI-native cloud orchestration.

Radiant's Radiant Cloud OS provides a unified platform from bare-metal provisioning through Kubernetes to higher-level AI services, allowing dynamic reshaping of infrastructure between training and inference workloads for enterprise customers.

The AI compute rental space is rapidly commoditizing as Meta enters against established neoclouds and hyperscalers, pressuring margins for pure-play compute providers and favoring vertically integrated players.

Alphabet earnings become bellwether for whether hundreds of billions in AI infrastructure spending converts to cloud revenue; backlog conversion speed is the key metric.

Major cloud providers show growth rate re-acceleration (AWS from 15-20% nadir to higher) as new models and applications drive unprecedented token consumption — with deep research reports now running hours/days instead of minutes — signaling sustained infrastructure demand that benefits both hyperscalers and neocloud specialists.

CoreWeave, Nebius, Lambda rent GPUs short-term; not platforms like AWS. Google explicitly calls them a 'bridge' until in-house capacity ready. Three exponentials on supply side (capex, chip perf, model efficiency) make pricing power impossible to forecast. Commodity boom/bust dynamics guarantee bust when hyperscalers pull back.

With 200+ neoclouds, lenders cannot assess operational quality (health checks, monitoring, managed Slurm/Kubernetes, SLA delivery). SemiAnalysis Cluster product provides hands-on testing and 150+ customer interviews to rank providers, enabling lenders to price execution risk and borrowers to negotiate contracts. Premium of 20-30% for reliable providers validates the differentiation.

Google Cloud's acceleration to $20B quarterly revenue at 32.9% operating margin proves the AI infrastructure buildout is generating profitable returns faster than peers, validating the $180B capex plan.

Deep technical collaboration between end-user, silicon vendor, and cloud provider — including Nvidia Aether for cross-environment Spark tuning — compresses GPU adoption timelines for complex production workloads, creating a template for enterprise AI infrastructure deployment.

Google Cloud's acceleration to 48% revenue growth with 30% operating margins demonstrates that AI infrastructure investments are translating into profitable cloud revenue, validating the hyperscaler model versus asset-heavy Neoclouds.

Microsoft is re-architecting Azure's data layer for GPU acceleration across all modalities — relational, vector, graph, streaming — because agentic workflows demand millisecond-level tool responses to maintain iteration velocity and token profitability.

Dean Ball's 'long compute, long neoclouds' thesis: even in an 'AI communism' scenario where models are free public goods, data center operators still earn margins on inference and training infrastructure.

Being cloud-native from day one lets digital banks continually re-architect and scale their tech stack, while traditional banks spend 5-10 years just migrating core systems to cloud, creating enduring technology advantage.

Neoclouds A and B both offering Nvidia H100s are undifferentiated — same hardware, same pricing, margin compression. Winners will blend 2-4 architectures: Nvidia for training/HPC, SambaNova for premium inference, AMD for cost-sensitive, custom for sovereign. This heterogeneity lets them offer tiered services (ultra-low latency, sovereign, high-throughput) at different price points, raising blended margins. SambaNova's partnership model (not building competing cloud) accelerates neocloud adoption.

New hyperscalers emerging: SpaceX AI (Colossus 1 → Anthropic), CoreWeave (Leopold holding), Fluid Stack (Anthropic deal). Anthropic's 10GW compute via investment deals (not capex) vs OpenAI Stargate unfunded. Compute access determines frontier model velocity. Neoclouds specializing in GPU clouds outpacing general clouds.

Agent workloads need low-latency, globally distributed execution with web access. Cloudflare's edge network + new scraper API + Workers gives infrastructure advantage; Vercel counters with developer experience. Winner captures platform economics for the agent economy.

Starlink's low-latency, high-bandwidth connectivity is displacing legacy aviation Wi-Fi (Viasat, Gogo). Private operators mandate it; United and American committed fleet-wide; Delta holdout risks customer loss. Production paused for next-gen dish with 2x bandwidth.

A new layer of inference-specialized clouds is emerging to provide the routing, harness, and memory abstraction that enterprises cannot build themselves, enabling hot-swapping between frontier and open models while preserving context — the infrastructure play for the model fungibility thesis.

xAI's data center build velocity (per Jensen) makes it the primary Blackwell debug partner for Nvidia, securing first-model advantage. This 'neocloud' model — vertically integrated model lab + compute — outperforms OpenAI's dependent model (paying margins to Microsoft/Oracle) and Anthropic's hybrid (TPU/Trainium + Nvidia).

Elon built Colossus faster than anyone, rented to Anthropic at $1.25B/month with 90-day cancellation; covers $12-19B capex in ~1 year. However, CoreWeave multiples value this business <$100B vs SpaceX's $2T+ ask — the gap is Elon premium and data-center-in-space optionality.

Meta (and SpaceX) pivoting from 'buy compute for proprietary models' to 'rent excess compute' — both rewarded by market (+10% Meta, SpaceX got 'extraordinarily high price'). Two endgames: (1) goldilocks — short-term rental, long-term proprietary model use; (2) cloud business is structurally great. Risk: if compute demand inflects, oversupply crashes economics. Rory: spend stops only when enterprise revenue growth (demand side) slows, not when supply-side capital runs out.

SpaceX and Meta entering the GPU rental market as new supply-side competitors. Jason: 'At the margin, the entrance of SpaceX and Meta into the Neocloud business... was worth exactly that 10 to 15% decline for Nebas and Kore.' Market share dilution for pure-play neoclouds unless demand grows faster than new supply.

GPU cloud providers are structurally supply-constrained, with 70% of fleets comprising older GPUs booked months ahead and contracts renewing early, signaling persistent inference demand.

Companies like CoreWeave that manage the full stack of GPU deployment — racking, power, cooling, engineering — capture value as AI labs outsource physical infrastructure to accelerate time-to-train.

Nvidia funds neo clouds to compete with hyperscalers, but neo clouds pay 70-80% GPU gross margins plus their own margin. Hyperscalers and Cerebras deploy own silicon at internal cost. Full-stack ownership (Google TPU, Cerebras wafer-scale) enables lowest token cost, though single-customer volume historically limited scale.

Specialized GPU cloud providers (neo clouds) like CoreWeave and Iron are Leopold's largest positions, providing turnkey AI infrastructure and benefiting from exponential compute demand.

Companies like CoreWeave, Applied Digital, CleanSpark, and Riot Platforms own the critical power licenses and grid interconnections that GPU manufacturers lack, allowing them to profit whether semiconductor stocks rise or fall.

Hyperscalers (Meta, Microsoft, Google) prefer opex compute rental over capex-heavy data center builds to smooth earnings and bypass 5-year permitting queues; neoclouds like Nebius, CoreWeave, and Iron that hold pre-approved sites, power contracts, and GPU deployment expertise are locking in $50B+ backlogs and becoming critical infrastructure partners.

Azure growth guiding down (37% from 40%) despite AI agent boom thesis is structurally concerning; if 20 agents 24/7 narrative were real, Azure should accelerate even at scale — deceleration suggests AI revenue is mostly OpenAI inference passthrough, not owned growth.

Nation-state data residency requirements (healthcare, defense, education) make public cloud obsolete; Hydra Host's OS lets any data center become a neo-cloud, with several billion in signed contracts across 60 data centers in two dozen countries.

Amazon's decade-long strategy — acquiring Annapurna Labs (2015), building Nitro, then Trainium — demonstrates that owning the full stack from chip to cloud service enables pricing and performance no merchant-silicon competitor can match. The Anthropic closed loop (equity + compute) compounds this advantage.

GPU customers rent entire clusters under long contracts, reducing the value of tenant isolation and CPU-era cloud architecture. Neoclouds can win through superior GPU performance and faster execution, though financing constraints and project failures will produce significant dispersion.