Outputmaxxing: Turning Compute Into Useful Work

Cluster waste compounds as organizations scale

An expensive cluster can leave accelerator cards unused even as its owners seek more compute. Opening the Latent Space conversation, Anj Midha argues that node allocation—the share of accelerator cards in use—should reach roughly 96%. He distinguishes it from MFU, another utilization measure that the exchange does not define, and estimates best-in-class MFU at 60–70%. These are his operating benchmarks, without specified hardware or workloads.

His diagnosis reaches beyond scheduling. The people funding a cluster, deploying it, operating it and measuring its output can be separated by a long chain of responsibilities. Small mismatches compound when rapid expansion replicates an imperfect operating model. Iterative cluster bringup—learning from each deployment before expanding—is his prescription. New AI capabilities make infrastructure discipline more valuable: waste costs more, and the margin for error shrinks.

Community benefits can support reliable supply

Midha makes alignment tangible with a proposal from Scott Nolan, introduced as General Matter’s founder: suppose compute costs $4 an hour. Charge $4.50 and transfer the additional revenue directly to the community hosting the data center. Midha says he would pay that premium at scale. Lower local electricity bills could serve the same purpose: give residents a visible stake in bringing the facility online.

The expected benefit extends to compute customers. Midha believes clear local benefits would strengthen community support and make future supply more reliable. This remains a proposed incentive design. His tentative estimate of the share of US data centers facing opposition comes with an immediate qualification that the figure may be overstated.

That preference for durable relationships also shapes procurement at AMP. Midha favors providers with long operating histories, credit records and experience surviving internet boom-and-bust cycles. He considers much of the novelty claimed for neoclouds marketing; established infrastructure operations still matter. He also predicts audits and regulatory investigations, without specifying when, as short-term decisions in the compute business face scrutiny.

Pool capacity without surrendering the baseline

Why should an independent infrastructure provider align incentives better than a company that owns the whole stack? Midha distinguishes two approaches to utilization: - Integration consolidates processes within one node. - Pooling separates a resource so multiple nodes can share it. AMP Grid is a pooling layer across clouds and chip types. It aims to make otherwise stranded compute available to other users, initially through scheduling and economic coordination. Making that capacity interchangeable across the stack remains an ambition.

His electricity analogy explains why demand patterns matter. A steel factory needs extra capacity at night; a shoe mill needs it during the day. Each keeps a guaranteed baseline while sharing capacity for non-overlapping peaks. Pooling works better when demand does not spike together. Midha positions AMP as an independent system operator—a neutral coordinator of suppliers and customers—and cites PJM as his historical model. In that arrangement, neoclouds supply capacity and venture firms sit on the demand side.

AMP began at scheduling, drawing on engineering leaders Seb and Mihai’s experience at Google. Discord supplies a useful comparison: it built its voice and video infrastructure internally while pooling communication demand across its users. The guest distinguishes his own developer-platform role from internal infrastructure, which he credits to Mark Smith. AMP takes pooling across organizational boundaries, giving labs such as Periodic a baseline and flexibility to change usage on shorter timelines.

The scheduling design separates two service expectations: - Guaranteed capacity supports base workloads. - Interruptible demand enters a queue where credits express priority; a higher-valued job can interrupt a lower-valued one. The guest connects this design to a16z’s Oxygen program and internal Google scheduling work. In his invented example, a job assigned ten credits takes priority over one assigned five. Credits provide a way to change priorities as research needs change.

The interviewer raises a consequential objection: internal bidding may not supply the commitment a strategic project needs. Leadership sometimes has to concentrate resources before a program’s value becomes legible through a marketplace. He attributes Google’s failure to anticipate GPT to this allocation problem, explicitly separating that interpretation from criticism of AMP.

Finance research that corporate priorities leave behind

AMP’s organizational answer includes Foundry, a capital business that incubates or invests in frontier labs, alongside its infrastructure business. The guest argues that researchers can advance a parent company’s frontier and later lose support when its priorities change. His hypothetical example is a shift from an omni model to coding. Independent financing gives the displaced research another route forward.

Publication creates a related problem. If labs withhold commercially promising work, public literature becomes a selected sample of what they know. The speakers describe this as commercial selection changing which research becomes public, with costs extending beyond the company. Their example of a DeepMind publication embargo is a secondhand, uncertain account; their broader case for independent labs is that valuable research should have a path to release.

AMP’s scale figures distinguish plans from available capacity: - Demand being secured: the guest first mentions approximately 1.3 gigawatts over four years, then 1.2 gigawatts and roughly $40 billion in cloud spend. He clarifies that the full amount is not secured and that current-year capacity has not been publicly confirmed. - Steady-state goal: a continuous 1.2-gigawatt baseline pool. - Estimated research spikes: roughly six gigawatts over four years, without specifying how that figure would translate into operational capacity. These figures do not establish deployed supply.

Better predictions must reach patient decisions

End-of-life prediction gives the infrastructure ambitions a concrete purpose. The guest recalls working with Stanford Medicine professor Nigam Shah using STRIDE, a longitudinal patient dataset whose access then required medical-school affiliation. His interest also draws on memories of funeral processions in Chennai, with dancing and drums celebrating reincarnation. He contrasts that experience with his perception of American medicine’s emphasis on postponing death.

A prognosis spanning six months to six years is too broad, he argues, to guide a patient’s choices. In his account, uncertainty and fear of malpractice encourage cautious recommendations; patients try more treatments, and hospitalization displaces time with family. The goal is more precise information that helps patients choose how to spend their remaining time.

The early experiments used simple neural networks, and the guest says regression models can also help with suitable data. Orders-of-magnitude better precision was the ambition, rather than a demonstrated comparative result: he gives no accuracy metrics or evaluation design. Liability remains a deployment bottleneck in his account because a model’s recommendation does not transfer responsibility away from the physician.

He describes returning to this concern through a new incubation after carrying it for 14 years. He links patient empowerment to data centers that benefit their host communities: one mission needs models, and training those models needs capacity. AMP was formed as a public benefit corporation, he says, and he invites people working in medicine, research and policy to contribute to both efforts.

Standardize the boundary where you delegate

The guest calls the connecting philosophy output maxing: pursuing the best outcomes from available resources. Belief in scaling does not justify wasting compute on a suboptimal approach. Maintaining too many architectures can also fragment effort and undermine standardization. He credits Anthropic’s development velocity partly to architectural concentration, while arguing that greater investment now makes exploration more affordable.

This turns alignment into an organizational engineering problem. Small teams coordinate through tight feedback loops; growth introduces specialization, delegation and interfaces that can lose information about the original goal. The proposed responses are to standardize communication through protocols and interface specifications, or develop a capability that adds enough capacity to change the constraint. A hypothetical room-temperature superconductor illustrates the second route; it is a speculative analogy, not a reported discovery.

The interviewer cites SF Compute’s attempt to standardize compute futures contracts, while saying he does not know its progress. The response identifies restricted visibility between parties, capital-market friction and operations as obstacles to starting an exchange. AMP’s proposed connection works in both directions: - An exchange with excess capacity receives demand from the grid. - An exchange with unmet demand receives supply from the grid.

The working implementation involves labs, universities and trusted parties; an open protocol remains the ambition. Availability can change quickly. The guest reports that expected year-end surplus disappeared over the preceding six weeks, leaving even founders who had raised billions asking for roughly 50 nodes within weeks. The node configuration is unspecified, but the anecdote illustrates the difficulty of obtaining capacity on short notice.

Alternative chips sharpen the standardization question. The guest recounts a startup adopting NVIDIA’s data-center reference architecture, preserving the rack footprint from an input/output perspective while focusing innovation on logic-die performance and systems co-design. External providers would handle hosting. The decision illustrates a practical rule: standardize the interface you delegate, then concentrate effort on your bottleneck.

The harder boundary concerns information. Designing chips and models together requires early knowledge of future model architectures. With a roughly two-year chip-development horizon in the guest’s account, the target workload can change before the hardware reaches market. An integrated organization can give both teams access to confidential plans; an independent startup loses that access. Trusted roadmap visibility becomes part of producing useful hardware, making relationships between labs and chip teams a capacity issue.

Research achievement can demonstrate leadership

The conversation turns to LM Arena when the interviewer calls the guest its founding CEO. The guest narrows that description to a five-month administrative stint recruiting engineering, product and design leadership while the founders completed their PhDs. He emphasizes that Anastasios had always been the company’s CEO. Anastasios’s research output and a side project already serving millions challenge the assumption that researcher and founder are mutually exclusive roles.

The guest’s investment argument separates ability from motivation: - Major research contributions require leadership: researchers must obtain resources, win organizational trust and get work released. - Running a company requires wanting the broader job: scientific conviction must extend to difficult decisions in hiring, management and customer relationships. Some researchers primarily want to publish.

Research support therefore need not depend on becoming a founder. The guest reports that AMP Grid donates excess compute and has set aside roughly a couple thousand H100s for support associated with nonprofits and university labs.

Trust also develops through ordinary consistency. The speakers recall an awkward first meeting at NeurIPS in New Orleans that warmed after discovering a shared connection to Singapore, followed by years of community interaction. The guest cautions against equating organizational scale with personal stress: influencing people can feel especially demanding to researchers without coaching or mentorship.

Technical precision reveals relationships that market categories hide. A question about real-time action prediction models is more useful than a broad question about world models. Researchers who understand their particular bottleneck may recognize people to learn from where investors see only competitors. The guest offers Pim’s reported enthusiasm for Black Forest Labs’ Flux work as an example of that camaraderie. Broad business categories can erase differences in missions and technical problems, introducing another kind of misalignment.

Preparation makes progress usable; culture sustains it

The guest distinguishes leading a field from winning a category: advance the state of the art and capture enough value to keep innovating while preserving the mission. He interprets Jensen and Dwarkesh’s podcast disagreement as a mismatch between an industry-layer model and a training-loop model. Different abstractions can conceal different assumptions, just as investors can mistake heuristics from an earlier era for timeless rules.

The interviewer asks whether Anthropic’s coding strength began with a lucky improvement that the company recognized and compounded. The respondent emphasizes preparation that makes useful data actionable: years of readiness allowed Anthropic to exploit context data and developer-provided context diffs, while early resource constraints forced efficiency. This is his organizational explanation for progress, rather than a disclosed training recipe.

The speakers then disagree about how durable culture can be: - Culture as a competitive advantage: the interviewer argues that team and culture underpin other advantages and let them compound. - Culture as a maintained practice: the respondent stresses that repeated actions must demonstrate the mission, or culture deteriorates. A system for making those tradeoffs consistently can become difficult to replicate, but it needs daily attention.

The guest describes Anthropic’s founding culture as confidence in scaling coupled with concern about unpredictable behavior and incomplete understanding of model internals. He anticipates Claude being used in consequential settings where errors could bring blame and liability. His defense of caution around early releases rests on uncertainty about those uses before deployment.

Repeated safety commitments create expectations others can enforce. Employees and the public can hold a company accountable when its actions drift from its stated mission. Founder leadership may help preserve that commitment through difficult tradeoffs, though the guest acknowledges that the advantage can become fragile.

Let the mission decide what to build—and what money is for

Scarcity, in the guest’s account, helped Anthropic define its P0, or top priority. He worries that abundant early funding can spare teams the tradeoffs that sharpen a culture. The build-versus-buy comparison makes the consequence concrete: - Anthropic treated coding as a central research objective: better coding could accelerate computer-based work and advance its mission of artificial general intelligence. That focus required declining other opportunities. - Periodic Labs treats coding as a tool: with superconductivity as its mission, it can use Claude and concentrate its research on physics.

Summarizing scientific literature and analyzing experimental data are different capabilities. The guest recalls evaluating frontier models while working in Stanford’s condensed matter physics group: they helped summarize papers but struggled with scientific data from the lab. The account gives no model names, protocol, scores or precise dates, so it describes that experience rather than current comparative performance. A claimed subsequent Periodic breakthrough likewise comes without technical results; the speakers propose a separate episode about the work.

The closing hiring example exposes a tension between mission and money. The guest recounts prospective Periodic hires accepting higher-paying jobs elsewhere, then seeking to return after the claimed breakthrough. He says they were refused, framing the decision around commitment and culture. The interviewer challenges the suggestion that the compensation differences were small: money can be life-changing to the person choosing the job, and many people have not been taught how to handle it.

The guest responds with memories of austere schooling at Rishi Valley and crowded scholarship housing in Singapore, where one roommate climbing into a bunk could shake another bed. He recalls enjoying those years and connects them to modest material needs.

The final reflection applies Goodhart’s law to wealth: when money changes from a byproduct of meaningful work into a measure of personal worth, the measure loses meaning. The exchange instead treats money as resources for pursuing a mission.