Tiger Global leads AI chip investing as inference and interconnect dominate

The concentration of capital into inference and optical interconnect startups signals that the industry's primary bottleneck has shifted from raw training power to data movement and serving costs. For the streaming and video ecosystem, this transition is critical as it dictates the unit economics of deploying generative AI for real-time video processing and personalized content delivery. As specialized architectures from companies like Groq and Etched move into customer deployment, the market is moving away from a general-purpose GPU monopoly toward a fragmented, task-specific hardware stack. Watch for whether Etched can convert its $1 billion in reported customer contracts into sustained revenue to justify its recent $21 billion valuation.

The AI chip investment landscape is intensifying as Cerebras moves toward a public listing. Cerebras filed for an IPO with a reported $10 billion contract from OpenAI forming a cornerstone of its growth narrative, alongside a key partnership with Amazon Web Services. The wafer-scale engine architecture offers a fundamentally different approach to Nvidia's GPU design, targeting massive parallelism with lower latency for inference workloads. If the IPO succeeds, it would provide Cerebras with public-market capital to scale production and potentially accelerate adoption among hyperscalers seeking alternatives to Nvidia's dominant position in AI compute.

The broader funding environment for inference and interconnect startups reflects a structural shift in where capital is flowing. Tiger Global's portfolio spans companies like Groq, Celestial AI, Lightmatter, Tenstorrent, Ayar Labs, and Etched, covering the full stack from inference processors to optical interconnect. Sovereign wealth funds including Temasek and Qatar Investment Authority have become increasingly active in late-stage rounds, while semiconductor incumbents such as Samsung and AMD are making strategic investments to secure supply chain positioning. This convergence of venture, sovereign, and corporate capital around inference-specific hardware marks a departure from the training-centric funding cycles of 2023 and 2024, when Nvidia's dominance made alternative architectures appear risky to institutional investors.

On the technical side, the inference hardware category is fragmenting rapidly by workload type. Groq's language processing units target transformer inference with deterministic latency, while Etched has raised significant capital around its Sohu chip designed specifically for transformer architectures. Optical interconnect companies like Lightmatter and Ayar Labs address the data-movement bottleneck that limits throughput in large-scale inference deployments. T-Mobile US has invested heavily in combining low-band, mid-band, and higher-frequency spectrum to balance coverage and performance for data-intensive applications, illustrating how network operators are preparing infrastructure for the AI workloads that these specialized chips will ultimately serve. For streaming video applications, the shift toward inference-optimized silicon directly affects the cost curve for real-time content processing, recommendation engines, and generative AI features that require sub-second response times at scale.