Anjney Midha

Anjney Midha is an AI investor and entrepreneur who founded AMP, a public benefit corporation building a compute grid for independent research teams. His work includes establishing Discord’s developer-platform organization and its Midjourney partnership, coleading a16z’s Oxygen GPU-access program with Guido Appenzeller, and helping incubate LMArena. In his June 2026 Latent Space interview, he described AMP’s goal as making “FLOPs flow like megawatts”: pooling compute so labs can scale without each having to provision for its own peak demand.[1][5][6][3][7]

Midha describes AMP as a horizontal, multi-cloud and multi-silicon coordination layer, beginning with scheduling and economic allocation. Members retain control of their baseload while sharing access to capacity for spikes. The rationale is that one lab’s training peak can coincide with another’s quieter research or deployment period, smoothing aggregate demand. He compares AMP’s intended role to an independent system operator in electricity markets, coordinating suppliers and users. AMP’s grid thesis also proposes shared automation for topology-aware scheduling, fast checkpointing, power management, and silent data corruption—problems individual labs otherwise tackle separately.[1][7]

Midha credits AMP engineering leaders Seb and Mihai with the Google scheduling experience behind its starting architecture. He connects the design to guaranteed capacity and interruptible jobs: teams keep capacity for routine workloads, while queued research jobs can yield to higher-priority demand. In June 2026, he described a target of 1.2 gigawatts of steady baseload and estimated a need for roughly six gigawatts of spike capacity over four years, explicitly saying that not all capacity had been secured.[1]

Midha calls his engineering philosophy “output maxing”—making more useful progress with available resources. He distinguishes node allocation, or how many cards are being used, from model FLOPs utilization, or how much theoretical compute becomes model work. His diagnosis is organizational as well as technical: rapid expansion and distance between funders, operators, and researchers can compound waste. He favors iterative infrastructure bring-ups and experienced data-center partners, and argues that community benefits should be part of reliable infrastructure procurement.[1]

Before AMP and his AI investing work at a16z, Midha studied bioinformatics at Stanford and left graduate study to join Kleiner Perkins, where he founded Edge. At its June 2015 launch, Edge combined seed investing with software development for portfolio founders. Roneil Rumburg and Ruby Lee joined to build tools for recruiting and operations; Midha said the team planned to spend about two days a week investing and the remainder engineering, designing, and shipping code.[4][14]

Midha subsequently cofounded and led Ubiquity6, which launched in 2017 with a consumer augmented-reality platform ambition. Its Displayland product let users scan physical environments with smartphone cameras. The company later pivoted to Backyard, a desktop platform for remote party games. Discord acquired Ubiquity6 in June 2021, with terms undisclosed; Backyard was discontinued as part of the announcement.[4][13]

At Discord, Midha became vice president of platform ecosystem, established its first dedicated developer-platform organization, and launched its Midjourney partnership. His remit was developer products. In the June 2026 interview, he credited Mark Smith with Discord’s internal infrastructure and explained that observing its in-house voice and video systems gave him context for a different architecture from AMP’s cross-provider pooling.[5][1]

Midha joined a16z as a general partner in July 2023; its later biography identifies him as a venture partner. By October 2024, he and Guido Appenzeller were coleading Oxygen. The program used portfolio-scale purchasing to help startups obtain GPU capacity with more flexible pricing and commitments. Midha explained the underlying risk: a young lab can plan a training run before it knows customer inference demand, making a long lease a bet on both future capacity needs and chip choices.[4][5][6]

His investment rationales also emphasize what developers can do with open models. With Matt Bornstein and Rajko Radovanovic, he coauthored a16z’s December 2023 Mistral announcement, arguing that community testing, fine-tuning, optimization, and red-teaming improve AI systems. With Marco Mascorro, Radovanovic, and Justine Moore, he coauthored the August 2024 Black Forest Labs announcement, advocating a division of work in which researchers build base visual models and application teams concentrate on workflows and user experiences.[12][11]

Midha helped LMArena move toward becoming a company. Latent Space’s January 2026 interview introduction credits him with forming an entity and providing grants before the researchers committed to a startup. Midha describes his own role as a five-month administrative stint, helping recruit engineering, product, and design leadership while Anastasios Angelopoulos and Wei-Lin Chiang finished their PhDs; he credits Angelopoulos with the company’s CEO role. In his May 2025 investment rationale, he favored continuous human-preference evaluation over static benchmarks and identified maintaining neutrality under commercial pressure as a substantial challenge.[3][1][8]

In September 2025, Midha reported working as a visiting scientist in Stanford’s Department of Applied Physics, evaluating frontier models on experimental data in condensed matter physics. He said their poor performance helped frame his conviction, shared with Periodic Labs founders Liam Fedus and Dogus Cubuk, that scientific AI needs feedback from physical experiments. His Periodic Labs investment announcement explained a loop of reading literature, running simulations, synthesizing materials, and learning from measured results. A16z led the company’s $300 million founding round.[9]

Midha’s interest in clinical prediction reaches further back. Recalling his Stanford bioinformatics work with Nigam Shah, he described experiments using simple neural networks to predict remaining life after a terminal diagnosis, with the aim of helping patients make better decisions about palliative care. In June 2026, he said he was developing a new incubation around more precise end-of-life prediction. He identified clinical liability as a continuing obstacle: improving a prediction does not by itself transfer responsibility for a wrong diagnosis from a physician to an AI system.[1]

Midha argues that accomplished researchers should be taken seriously as potential CEOs: winning resources, earning institutional trust, and producing leading research already require leadership. He also draws a distinction between wanting to publish and wanting to run a company. In his view, the latter requires defending decisions across hiring, customers, and the team. He treats culture as fragile rather than permanent, sustained by daily choices, and warns that abundant early funding can let a lab postpone defining its highest priority.[1]

In a September 2026 discussion with OpenRouter co-founder Alex Atallah, the publisher describes Midha’s concern that token fraud could grow as AI usage becomes more valuable and that autonomous agents could amplify attacks. The interview gives a concrete example of a free image-generation trial being resold through fraudulent accounts. The prediction about future agent-driven fraud is a view about a possible risk, not evidence that the forecast has already happened.[15]