The shared language of AI investing

AI systems that perceive, reason about, and act on the physical world through sensors and actuators, spanning robots, autonomous vehicles, and smart machines. It extends generative AI from digital outputs to physical actions, and is the framing NVIDIA and others use for the next wave of AI value creation.

AI systems that autonomously plan multi-step tasks, call tools, and take actions toward a goal with limited human oversight, rather than just responding to single prompts. Agentic workflows became a dominant 2025-2026 enterprise adoption theme and a core investment thesis for AI application startups.

AI that learns and operates through a physical body interacting with real environments, coupling perception, control, and learning in robots and humanoids. It is seen as a key path toward general-purpose robotics and a major hard-tech investment frontier.

A large model pre-trained on broad data at scale that can be adapted to many downstream tasks via fine-tuning or prompting, such as large language and multimodal models. Foundation-model labs command the largest AI funding rounds and anchor the industry value chain.

Stacked DRAM connected to processors via a wide interface to deliver very high memory bandwidth, essential for feeding data-hungry AI accelerators. HBM3E and emerging HBM4 are supply-constrained bottlenecks in the AI hardware chain, dominated by SK Hynix, Samsung, and Micron.

The unit economics of AI where cost, pricing, and margins are measured per token processed for input and output. Falling cost-per-token from better chips and models reshapes the economics of AI applications and is a core diligence metric for investors.

The phase where a trained AI model runs on new inputs to generate outputs, as opposed to training. As deployment scales, inference is becoming the larger and recurring share of AI compute demand and cost, driving investment in inference-optimized chips.

A nation's capacity to build and control its own AI compute, data, and models using domestic infrastructure, driven by security, economic, and regulatory motives. Sovereign AI has become a major source of government-backed data-center and chip demand worldwide.

A fully autonomous ride-hailing service operating without a human driver, offered by players such as Waymo and China's Baidu Apollo Go and Pony.ai. By 2026 robotaxis operate commercially in multiple cities and are a closely watched autonomy-monetization milestone.

A robot with a human-like body designed to operate in human environments and use human tools, from Tesla Optimus to Figure and Chinese entrants like Unitree. Humanoids are a flagship embodied-AI category attracting heavy venture and strategic investment in 2025-2026.

A training method that aligns model behavior with human preferences by fine-tuning on a reward model learned from human comparisons of outputs. RLHF was central to making chatbots helpful and safe, and is increasingly complemented by RLAIF and reinforcement learning on verifiable rewards.

A model architecture that routes each input to a small subset of specialized expert sub-networks, so only part of the parameters activate per token. MoE lets models scale total capacity while keeping inference cost lower, and underpins many frontier 2025-2026 models.

The private-market ecosystem for investing in unlisted AI companies through venture and growth rounds before any public listing. In 2025-2026 it is characterized by mega-rounds into foundation-model and AI-infra firms at record valuations.

Late-stage private investment in a company expected to go public relatively soon, typically at a large valuation with lower expected multiples but shorter time-to-liquidity. Pre-IPO AI names are a focus for crossover and growth investors seeking exposure ahead of listings.

A fund manager operating both a US-dollar fund and an RMB fund in parallel, common among China-focused VC/PE firms to serve different LPs and portfolio needs. The USD and RMB vehicles differ in investors, exit paths, and regulatory regimes.

A financing round priced at a lower per-share valuation than the company's previous round, diluting existing holders and often triggering anti-dilution adjustments. Down rounds rose after the 2022-2023 reset and remain a risk for overvalued AI and tech startups.

A battery that replaces the liquid electrolyte with a solid one, promising higher energy density, faster charging, and improved safety over conventional lithium-ion. As of 2026 it is nearing early automotive commercialization and is a closely tracked hard-tech and EV investment theme.

A wide-bandgap semiconductor material that handles higher voltages, temperatures, and switching frequencies than silicon, improving efficiency in EV inverters, chargers, and power systems. SiC is a strategic power-electronics segment led by firms like Wolfspeed, Infineon, and STMicroelectronics.

NVIDIA's next-generation AI GPU platform succeeding the Blackwell generation, named after astronomer Vera Rubin and slated to ramp around 2026. It pairs new Rubin GPUs with Vera CPUs and next-gen HBM to further scale AI training and inference performance.

An open protocol introduced by Anthropic in late 2024 that standardizes how AI applications connect models to external tools, data sources, and services. MCP has become a widely adopted interoperability layer for building agentic AI systems.

The practice of structuring and publishing content so it is accurately retrieved, cited, and surfaced by AI answer engines like ChatGPT, Perplexity, and Google's AI Overviews. GEO extends SEO for an era where AI-generated answers, not blue links, mediate discovery.

The large capital expenditure that hyperscalers and cloud providers spend building AI data centers, including GPUs, networking, power, and cooling. AI-driven data-center capex reached hundreds of billions of dollars annually by 2026 and is a key macro driver for the AI hardware supply chain.

Empirical relationships showing that AI model performance improves predictably as compute, data, and parameters scale up. In 2025-2026 the focus broadened from pre-training scaling to inference-time (test-time) compute scaling for reasoning models.

A class of models that spend additional inference-time compute to think step by step before answering, improving performance on math, coding, and complex problems. Pioneered by OpenAI's o-series and followed by others, reasoning models reshaped the frontier in 2025-2026.

A technology that integrates optical engines directly alongside switch or accelerator chips in the same package to cut power and boost interconnect bandwidth for AI data centers. CPO is an emerging bottleneck-relief technology gaining investment attention as AI networking scales.

In a fund, Limited Partners are the investors who provide capital, while the General Partner manages the fund and makes investment decisions, earning management fees and carried interest. The LP-GP structure is the standard governance model for venture capital and private equity.