Capital Shifts to AI Foundation Models and Robotics: OpenAI and Embodied Intelligence Emerge as New Anchors
Capital Shifts Toward “Root Technologies”: As OpenAI Stake Hits 30%, the Primary Market Is Redrawing the AI Power Map
Recently, Eightco added a $40 million investment in OpenAI, bringing its total stake to $90 million—representing 30% of the fund’s current portfolio. This figure far exceeds conventional venture capital allocation limits for any single investment. Almost simultaneously, multiple purchase requests appeared in 36Kr’s “Capital Intelligence Bulletin” (Issue #181): “Seeking pre-IPO shares in Anthropic,” “Seeking pre-IPO shares in CloudDeep, UBTECH, DaTuo, and Star Dynamics”—all robotics companies. On the surface, these appear to be isolated secondary-market share transactions; in reality, they constitute an exceptionally rare “signal resonance” across the primary market: capital is collectively exiting the noisy application layer and instead doubling down on two foundational pillars of the AI era—control over foundational models and execution authority in embodied intelligence. This is not a mere style shift—it is a paradigm shift. The AI race has moved beyond “whose app looks flashier” into a hard-core phase defined by three critical questions: Who controls model licensing gateways? Who holds compute orchestration rights? And who delivers mass-produced robots first?
From “Application Hype” to “Root Infrastructure Positioning”: Why Models and Robots Have Become the New Anchors
Over the past two years, massive capital flooded AI-native applications—writing assistants, design tools, educational plug-ins, and more. Yet several recent high-engagement posts on Hacker News reveal the application layer’s structural fragility. One post titled “Show HN: I made an email app inspired by Arc browser” garnered over 1,000 upvotes—but no one asked whether its underlying inference relied on a GPT-4 Turbo license. Another, “OpenCode – The open source AI coding agent,” is open-source in name, yet its core model still depends on closed API calls. Most metaphorically revealing was “France’s aircraft carrier located in real time by Le Monde through fitness app,” where Le Monde pinpointed the location of France’s Charles de Gaulle aircraft carrier using Strava’s public heatmap data. These cases expose a fatal weakness of the application layer: extreme dependence on external data sources and model services, with little technical depth of its own—and thus highly vulnerable to upstream policy shifts or commercial terms that cut off supply. When OpenAI adjusts its API rate limits—or Anthropic tightens distribution terms for its Claude Enterprise edition—countless applications instantly stall. Capital has awakened to a stark truth: within the AI stack, the application layer is merely a “replaceable skin”; only the model layer and execution layer constitute the “irreplaceable skeleton and muscle.”
Eightco’s allocation of 30% of its portfolio to OpenAI reflects no bet on valuation bubbles—but rather a wager on the scarcity of model control rights. Though unlisted, OpenAI’s models have become the de facto industry-standard interface: developers default to gpt-4o as their prompt-engineering benchmark; enterprises evaluating AI services ask first, “Is it compatible with the OpenAI ecosystem?” This control directly translates into commercial moats: third-party tracking shows OpenAI’s API accounted for 68% of global enterprise-grade AI API call volume in Q1 2024—far outpacing Anthropic (12%) and Google (9%). Holding OpenAI shares is effectively holding a “digital access card” to the entire AI application ecosystem. Likewise, the intense demand for pre-IPO Anthropic shares on 36Kr’s bulletin board isn’t about chasing valuation—it’s about securing a seat at the table for shaping constitutional AI governance frameworks. Anthropic’s Constitutional AI principles have already been cited repeatedly in draft versions of the EU AI Act; owning its shares may grant future influence over the global pace and substance of AI compliance standard-setting.
Embodied Intelligence: A Quantum Leap from “Lab Curiosities” to “M&A Battleground”
If the model-layer contest is for “sovereignty of the brain,” the surge in demand for robotics pre-IPO shares targets “sovereignty of the body.” Notably, all companies listed in 36Kr’s purchase requests—CloudDeep (quadruped robots), UBTECH (humanoid robots), DaTuo (cloud-based robotic OS), and Star Dynamics (general-purpose robot platforms)—focus squarely on mass-producible physical robots, not just algorithms or simulation environments. This signals a fundamental shift in industrial logic: before 2023, robotics investments prioritized “demo wow-factor” (e.g., backflips, bottle-opening); starting in 2024, capital rigorously scrutinizes “ramp-up curves for mass production”—including unit-cost decline slopes, monthly production capacity growth rates, and failure-rate decay cycles. The Hacker News post “Baltic shadow fleet tracker – live AIS, cable proximity alerts” ignited broad discussion precisely because it fused AIS vessel data streams with geofencing algorithms to achieve millisecond-level dynamic control over physical assets (cargo ships). This foreshadows a pivotal truth: the ultimate realization of AI value must anchor itself in physically executable units within the real world.
The capital markets have already cast their vote—in hard currency. In UBTECH’s 2023 Hong Kong IPO, 73% of proceeds were explicitly earmarked for “humanoid robot mass-production line construction.” CloudDeep completed its Series B+ round in 2024, led by a national-level manufacturing industry fund whose terms included “first-right-of-refusal procurement” and “joint development of next-generation industrial inspection robots.” Even as Eightco doubles down on OpenAI, several of its limited partners—industrial conglomerates—are accelerating direct engagement with precisely these robotics firms. Models provide decision-making intelligence; robots deliver physical execution. Only their tight coupling yields closed-loop commercial value. This explains the urgency behind pre-IPO share acquisitions: the primary market has lost patience waiting for the next funding round—it seeks instead to secure immediate access to existing production capacity and supply-chain control, seizing a critical time window ahead of imminent M&A consolidation.
The Next Phase of Competition: Tripartite Positioning in Model Licensing, Compute Access, and Mass Delivery
Based on current signals, the infrastructure war for AI will unfold across three decisive fronts:
First, the battle for model licensing rights is intensifying. OpenAI has launched its “Enterprise API Tier” tiered licensing program, granting top-tier customers custom fine-tuning permissions and dedicated inference clusters. Anthropic, meanwhile, rolled out its “Constitutional Partner Program,” requiring signatories to commit formally to AI ethics implementation before gaining deep access to Claude 3.5. Capital’s rush to acquire pre-IPO shares is fundamentally a play to secure a seat at the future licensing negotiation table—shareholders will gain priority access to model fine-tuning interfaces, latency SLAs, and co-branded marketing rights.
Second, compute access rights are emerging as a new moat. Model performance leaps hinge critically on compute density. Eightco disclosed that 15% of its OpenAI investment is specifically allocated toward co-building an “edge-cloud collaborative inference network,” targeting average customer inference latency under 200ms by end-2024. This signals a strategic pivot: competition among model service providers will increasingly center not on “parameter count,” but on “compute orchestration efficiency.” Part of the premium attached to robotics pre-IPO shares stems from their proprietary GPU clusters—already operating at 7×24 full utilization. Such infrastructure is far harder to replicate than algorithmic patents.
Third, mass-delivery capability determines M&A leverage. Industry consensus is crystallizing: humanoid robots priced below RMB 150,000 per unit represent the threshold for large-scale commercial viability. Currently, only UBTECH and Star Dynamics have publicly announced entry into this price band. Capital’s pre-IPO share purchases are, in essence, advance reservations for production quotas beginning Q3 2024—whoever secures first-mover access to ten-thousand-unit delivery capacity will hold the initiative in defining industry standards.
When Eightco’s $90 million commitment and 36Kr’s dozens of urgent purchase requests converge on the same timeline, what we witness is not capital frenzy—but rational reconstruction. The next act of AI belongs not to “disruptive applications” showcased in glossy pitch decks, but to pragmatic builders who quietly fortify root-model technologies, relentlessly optimize robot yield rates, and willingly pay a premium for compute infrastructure. Infrastructure wars produce no fireworks—only the whirring of cooling fans on AI chips and the low, steady hum of robotic arms operating continuously in factories. That is the true background score determining the AI power map for the next decade.