The $200M Bet: Why AI Networking Is the New Infrastructure Bottleneck
\n\nEridu just raised over $200 million in Series A funding—one of the largest early-stage rounds in recent memory—to build networking hardware for AI data centers. Not software. Not a new model architecture. Hardware that moves data between GPUs.
\n\nIf that sounds boring, you're missing the story. This is where AI's next major constraint lives, and the market is pricing that realization at nine figures.
\n\nThe Bottleneck Migration
\n\nEvery generation of computing infrastructure hits a different wall:
\n\n1980s-1990s: CPU speed was the constraint. Faster processors unlocked new capabilities. Moore's Law dominated strategy.\n\n2000s-2010s: Memory became the bottleneck. RAM speed and capacity determined what you could run. The rise of in-memory databases (Redis, Memcached) and memory-optimized instances reflected this shift.\n\n2010s-2020s: Storage I/O took center stage. SSDs replaced spinning disks. NVMe emerged. Database performance became primarily a storage throughput problem.\n\n2020s-now: Network interconnects are the new constraint.\n\nAI training runs on thousands—sometimes hundreds of thousands—of GPUs operating in parallel. These GPUs need to constantly synchronize gradients, share parameters, and communicate intermediate results. Traditional data center networks were never designed for this communication pattern.
\n\nThe result: your $50,000 H100 GPUs sit idle waiting for data transfers to complete. You're burning electricity and capital on compute that spends a non-trivial percentage of its time blocked on network I/O.
\n\nThis is the problem Eridu is solving.
\n\nWhy Traditional Data Center Networks Don't Work for AI
\n\nStandard data center architectures use Ethernet switches in a hierarchical topology: top-of-rack switches connect to aggregation switches, which connect to core routers. This design works well for typical cloud workloads where most traffic flows north-south (client to server) or between loosely-coupled services.
\n\nAI training has a fundamentally different traffic pattern:
\n\nAll-to-all communication. In distributed training, every GPU needs to talk to every other GPU. Gradients computed on GPU #1 need to reach GPU #47,382. This creates a full mesh of communication at massive scale.\n\nLatency sensitivity. A single slow transfer blocks the entire training step. The 99th percentile matters as much as the median. Traditional networks optimize for throughput; AI needs consistent, predictable low latency.\n\nMassive bandwidth requirements. GPUs can compute faster than networks can move data. A single H100 can perform 2 petaFLOPS of computation but is bottlenecked by a 3.2 Tbps interconnect. Multiply that across thousands of GPUs and you need data center networks with aggregate bandwidth measured in petabits per second.\n\nPower constraints. Traditional switches consume significant power relative to their throughput. When you're running AI clusters at the gigawatt scale (see: Thinking Machines' recent 1GW Nvidia partnership), network power becomes a meaningful fraction of total infrastructure cost.\n\nThe math breaks. A conventional data center network designed to handle typical cloud workloads crumbles when you try to coordinate 100,000 GPUs training a foundation model.
\n\nWhat Eridu Is Building
\n\nWhile Eridu remains in stealth and hasn't publicly disclosed technical details, the problem space suggests several likely innovation vectors:
\n\n1. Optical Interconnects
\nElectrical signaling over copper hits physics limits at high speeds. Optical interconnects can achieve dramatically higher bandwidth with lower latency and power consumption. Co-packaged optics (where the optical components sit directly on the switch chip) eliminate conversion latency and reduce power by 30-50%.
\n\n2. Custom Switching ASICs
\nGeneral-purpose Ethernet switches optimize for flexibility. AI workloads are predictable and repetitive. A custom ASIC designed specifically for GPU-to-GPU communication can strip out unnecessary features, reduce latency by microseconds, and improve power efficiency by 2-3x.
\n\n3. Novel Topologies
\nTraditional hierarchical networks create bottlenecks at aggregation layers. Fat-tree, Clos, or dragonfly topologies provide multiple paths between any two nodes, reducing congestion. Even more exotic designs—like 3D torus networks or reconfigurable optical circuit switching—may eliminate traditional switch hops entirely.
\n\n4. In-Network Compute
\nOffloading gradient aggregation or reduction operations to network switches can reduce the data that needs to travel. If switches can perform partial computations on data in flight, you reduce both bandwidth and latency.
\n\nThe exact approach matters less than the core thesis: AI workloads justify purpose-built networking hardware.
\n\nWhy $200M+ Makes Sense
\n\nA $200 million Series A is unusual. It signals several realities about this market:
\n\nCapital-Intensive Hardware Development
\nBuilding custom ASICs costs tens of millions per generation. You need fab partnerships (TSMC, Samsung), extensive validation, multiple silicon spins. Optical components require precision manufacturing. This isn't software with marginal costs near zero—hardware development requires serious upfront capital.
\n\nWinner-Take-Most Market Dynamics
\nData center infrastructure tends toward standardization. If Eridu's technology becomes the de facto standard for AI networking, they capture the entire market. Investors are betting on a potential Nvidia-style position in the network layer.
\n\nMassive TAM
\nAI infrastructure spending is projected to exceed $500 billion annually by 2030. Networking represents 10-15% of total data center capex. Even a modest market share in AI data center networking translates to billions in revenue.
\n\nDefensibility
\nCustom silicon creates multi-year moats. Once a hyperscaler deploys your hardware in production, switching costs are enormous. Software can be replaced; ripping out and replacing physical network infrastructure across a multi-gigawatt data center is a $100M+ decision with months of downtime.
\n\nFrom an investor perspective, this is infrastructure that every major AI lab needs. OpenAI, Anthropic, Google, Meta, Microsoft, xAI—all run into this bottleneck. The market isn't speculative; the pain is acute and immediate.
\n\nWhat This Means for You
\n\nIf you're building AI products, this matters:
\n\nTraining Costs Will Drop
\nNetwork bottlenecks mean you're paying for idle GPUs. Better interconnects improve GPU utilization from ~70% to 90%+, directly reducing training costs. A 20-point utilization gain on a $10M training run saves $2M.
\n\nLarger Models Become Feasible
\nThe viable size of models you can train is constrained by how well you can parallelize across GPUs. Better networking enables larger-scale parallelism, which enables larger models. We may see 10T+ parameter models become economically trainable in the next 24 months purely from infrastructure improvements.
\n\nInference Latency Improves
\nWhile training gets the headlines, inference clusters also benefit. Multi-node inference (where a single request spans multiple GPUs) becomes faster. Speculative decoding, mixture-of-experts routing, and other advanced inference techniques require low-latency inter-GPU communication.
\n\nCompetitive Advantage Shifts
\nToday, access to GPUs is the constraint. Tomorrow, it's network architecture. Companies with superior interconnect infrastructure will train faster and cheaper. This isn't a minor edge—it's the difference between iterating weekly vs. monthly.
\n\nThe Broader Pattern: Infrastructure Specialization
\n\nEridu is part of a larger trend: general-purpose infrastructure is being replaced by AI-native alternatives.
\n\nGPU clusters replaced CPU-based HPC. Nvidia didn't just make faster chips; they reimagined compute architecture for parallel workloads.\n\nCustom training frameworks replaced general ML libraries. PyTorch and JAX are designed for deep learning, not generic scientific computing.\n\nVector databases replaced relational databases for embeddings. Pinecone, Weaviate, and Chroma optimize for approximate nearest neighbor search, not ACID transactions.\n\nNow: AI-native networking replaces Ethernet. The same pattern. General-purpose infrastructure works until scale demands specialization.\n\nThis creates an entire category of infrastructure companies—each solving one piece of the AI stack with purpose-built tools. Whoever owns these layers owns the economics of AI.
\n\nThe Risk: Premature Specialization
\n\nNot every bet on specialized infrastructure pays off. Google's TPUs delivered impressive performance gains for specific workloads but struggled to achieve broad adoption outside Google. Many AI-specific chips (Graphcore, Cerebras, SambaNova) raised hundreds of millions but failed to displace Nvidia.
\n\nEridu faces similar risks:
\n\nAdoption inertia. Hyperscalers have existing Ethernet infrastructure and operational expertise. Ripping it out for a proprietary solution requires overwhelming performance gains.\n\nStandardization battles. If multiple vendors build incompatible AI networking solutions, the market fragments. Nobody wants vendor lock-in on core infrastructure.\n\nNvidia's moat. Nvidia owns the full stack from GPU to networking (NVLink, InfiniBand via Mellanox acquisition). Eridu needs to offer 10x better economics to convince customers to abandon an integrated solution.\n\nWorkload evolution. If AI architectures shift away from massive clusters toward smaller, edge-deployed models, the demand for hyperscale networking diminishes.\n\nThat said, investors clearly believe the pain is acute enough—and the TAM large enough—to justify the risk. $200M+ in Series A funding doesn't happen on speculative bets. The hyperscalers are likely already testing early prototypes.
\n\nWhat to Watch
\n\nThis space will move fast. Here's what signals progress:
\n\nPublic deployments. When OpenAI, Anthropic, or Meta announce using Eridu hardware in production, adoption accelerates rapidly. Credibility in this market comes from production workloads, not benchmarks.\n\nInteroperability standards. If Eridu's technology gets adopted into an open standard (like PCIe, NVLink, or CXL), network effects kick in. Proprietary solutions struggle; standardized ones compound.\n\nPower efficiency metrics. Watts per terabit becomes a critical KPI. If Eridu achieves 50% power reduction per unit of throughput, the economics become undeniable at gigawatt scale.\n\nSecondary market activity. Who else enters this space? If Intel, Broadcom, or Marvell announce AI-native networking ASICs, it validates the category but intensifies competition.\n\nIntegration partnerships. Does Eridu partner with server OEMs (Dell, HPE, Supermicro) or cloud providers (AWS, Azure, GCP) to offer turnkey solutions? Distribution matters as much as technology.\n\nThe Bottom Line
\n\nAI's next bottleneck isn't compute. It's not memory. It's the network.
\n\nEridu's $200M Series A is a bet that purpose-built networking hardware will become as essential to AI infrastructure as GPUs themselves. If they're right, this becomes a multi-billion-dollar category with winner-take-most dynamics.
\n\nFor companies building AI systems, this means:
\nThe infrastructure layer of AI is still being built. Compute got solved first (Nvidia). Storage is being solved now (vector databases, object stores). Networking is next.
\n\nThe winners in this layer won't just enable AI—they'll determine who can afford to build it.
\n