Did Turbopuffer write the new template for infra startups?

All roads lead to Montreal. For the past few months, almost every conversation I have about how to scale an AI-native infrastructure company ends up revolving around the same question. How did a tiny, unassuming infrastructure startup with almost no venture backing quietly manage to break every single rule of the modern scaling playbook and build an absolute giant. That company is Turbopuffer, founded by Simon Eiskildsen, a Danish entrepreneur based in Montreal.

If you spent 2023 and 2024 watching the AI infrastructure gold rush, you know the script. Raise a massive $50M Series A on a slide deck, hire a bloated sales team, burn through cash to win developer mindshare, and spin up an army of developer advocates to flood the web. Tack on some expensive enterprise salespeople on top of that, and you are good to go. You’re burning a ton, but that’s what it takes to scale, right? Wrong.

But first, what is Turbopuffer? Turbopuffer is a serverless, developer-first vector and full-text search database engineered specifically for scale-heavy AI applications. It fundamentally beats traditional vector databases by abandoning expensive, always-on in-memory (RAM) setups in favor of a decoupled architecture that queries data directly from cheap object storage like Amazon S3. By shifting the heavy lifting away from RAM, Turbopuffer allows companies to slash their infrastructure and scaling costs by up to 95% without sacrificing the snappy, low-latency performance required for real-time AI retrieval and search. 

The story of Turbopuffer is a total narrative violation of that entire worldview. Turbopuffer launched publicly in late 2023. By March 2026—just 19 months after hitting their first million—they crossed a $100 million ARR run-rate while operating fully profitably. The most provocative part of this quest isn't just the speed of their growth, but their radical capital efficiency. They reached this incredible scale with less than $1 million in total funding raised and a hyper-focused team of only about 38 employees.

Anomaly or playbook? Is Turbopuffer’s staggering, capital-efficient growth trajectory anomaly or does it offer some sort of emerging template for AI-native infrastructure companies? Is there a secret sauce or are there lessons here to be unpacked and applied to other companies?

Over the past few weeks, I’ve talked to a lot of people to try to figure out what is powering Turbopuffer’s scale. I’d love to have talked to someone at Turbopuffer (believe me, I reached out to many of them), but there are apparently too few of them and they are too busy kicking ass and taking names. But even without their direct help I think I’ve been able to distill a six-part framework that captures much of their success formula. I’m sharing it here for two reasons: first, maybe it will be helpful to others. but most importantly, maybe i’m missing something and I’d love to learn more about how this worked. And if any of you have a friend at Turbopuffer itself, please hook me up!

1. Small team of empowered people. 

The only way to grow from zero to $100M with under 40 people is to make sure those people are highly empowered A players. In the zero-interest-rate policy (ZIRP) era, headcount was treated as a proxy for success. In the AI-native era, that’s gone. Advanced AI tooling makes engineers and GTM people more effective than ever - and smaller teams appear to be the way of the future. The days of startups bragging about headcount are over. Highly motivated crack teams are the only way forward. 

Ask yourself: Do I really need more people? Is everyone on the team an A player? Do people have the motivation and authority to drive the business forward? Are we using modern AI tooling effectively to minimize headcount and reduce management cycles?

2. Direct authentic communication with clear storytelling

Turbopuffer’s website is a work of genius. It speaks clearly and compellingly to the target audience and effectively walks through the entire sales process. The ROI calculator makes it very easy to understand what benefits Turbopuffer will bring in a specific use case. The writing is authentic and straightforward. The design is grounded and serious. The entire experience makes it clear that the target audience is the developer customer and not the San Francisco VC hype factory. The same is true for all of Turbopuffer’s communications. In an era of flashy launch videos, over-designed landing pages, and mountains of AI slop content, Turbopuffers stands as an example of straight-forward, honest communication that is perfectly pitched at the target audience and no one else.

Ask yourself: How fast can your target audience understand what you are offering them and how it makes their lives better? Does your company speak in an authentic voice?

3. Focus on "kingmaker" reference customers first

Especially in this era of agentic AI, deep human trust is essential and scarce. Engineers trust what other elite, highly respected engineers use. Turbopuffer landed early customers like Cursor, Notion, Linear, and Anthropic. These early customer wins immediately positioned Turbopuffer as both trusted and aspirational. They also positioned Turbopuffer to naturally benefit from the underlying growth of these customers through a consumption pricing model. Some of this may have been luck, but I suspect a lot of it was deliberate. These flagship accounts make the company's geographic location (distribution and not in San Francisco) even more striking. Being physically present in San Francisco may make many things easier, but the real goal is winning the customers and mindshare the matters. Turbopuffer never seems to have lost sight of that goal.

Ask yourself: Who are the lighthouse accounts in our space? Can we expand that into an ICP definition that can help us identify more? Does our feature set map to the needs of that ICP? What shortcuts or hacks can we implement to accelerate adoption at the reference accounts that will accelerate downstream growth?

4.  Position as a drop-in replacement for something people were already doing.

Part of the reason Turbopuffer was able to scale so quickly was built into the nature of the product itself. Turbopuffer was built as a drop-in replacement for workflows that engineering teams are already running. Turbopuffer didn't ask companies to invent a new paradigm; they targeted organizations already struggling to scale vector search on Postgres, using expensive alternatives like Pipecone, or bleeding cash trying to host their own memory-heavy index clusters. Many companies struggle to explain why a customer should try a new thing. Turbopuffer doesn’t. Turbopuffer simply offers a way to do the thing you are already doing in a much better way. This is far easier to write up as a strategy point than to implement in real life. 

Ask yourself: Does our product represent an easier/better/faster/cheaper way of doing something our customers are already doing? If not - can we reframe our product to be easier to understand or adopt? Can we reference off an existing process or product?

5. Minimal integration surface area

Another key feature of the Turbopuffer product that unlocked fast growth is the minimal integration surface area. Not only is the Turbopuffer a simple drop-in replacement for an existing process, but it’s an almost trivially easy installation. VectorDBs exist distinct from other enterprise systems and other databases. This reduces the amount of buy-in required at a customer. It also reduces the cost and time required to test the product. The trick here is scoping the product so that it’s as easy to adopt as possible. Sometimes this means making the product bigger so that the integration surface area is minimized. Often, it can mean makes the product smaller. But either way, the goal is reduced integration friction. Do not build a bloated, "do-everything" multi-model product from day one. Limit the scope of your initial integration so the distance between a developer's curiosity and a paying, production workload is as short as possible.

Ask yourself: Assuming a customer wants your product, how fast and easily can they implement it? What barriers are in place that impede quick adoption? What needs to be removed from or added to your product to achieve the fastest possible time to value?

6. One-dimensional competitor comparison

Look again at the Turbopuffer website. It leads with a cost calculator. The potential customer inputs their configuration, and the calculator spits out a cost. Turbopuffer knows If your value proposition requires a complex, multi-variable spreadsheet to prove ROI to a CFO, the sale will drag on forever. But they also know that they are cheaper than the alternatives by 1-2 orders of magnitude. Their website frames up the relevant comparison and provides it directly to the potential customer. Turbopuffer limited their value metrics to one or two undeniable, easily measured KPIs: Cost (how much you pay per million vectors) and Latency (query speed). Because they engineered an object-storage-native engine, they routinely proved a 10x to 100x cost reduction while maintaining sub-10ms warm query latency. This sets up the success criteria for any trial and speeds the path to a win. 

Ask yourself: How do you want your customers to assess your product? What are the 1-2 KPIs they will use to convince themselves (and their organizations) to test, adopt, and purchase your product? Are you promising many vague benefits or 1-2 clear quantifiable benefits?

Is this template replicable?

Not every product can benefit from all aspects of the Turbopuffer playbook - and there are other playbooks that are also working well. Supabase’s partnership with Lovable is another great example. What’s interesting to observe about the Turbopuffer playbook is that while some of the elements (1-3) can and should be applied by almost any infrastructure company, some (4-6) require thinking very critically about the essence of your product. Turbopuffer’s success is about more than brilliant engineering and smart GTM strategies. It’s also about recognizing that product strategy and GTM strategy are - increasingly - one and the same. There are aspects of the product itself that are directly linked to the company's ability to generate rapid growth. Not every product should be altered to enable these strategies, but if you are building in infrastructure software today, you should at least be stress-testing that possibility. Maybe - just maybe - there are growth unlocks available to you in your product strategy itself.

If you think I’ve missed an aspect of the Turbopuffer growth story - especially one that might help other founders scale - please let me know!

Gil Dibner

Could the future of software be fluid
How do we get the best of AI without losing the soul of software?

The future belongs to young missionary teams
Why it makes more sense betting on youth in the current moment

The AI-native enterprise playbook
Ten real-time observations on a rapidly evolving playing field

No more painting by numbers
It’s the end of the “SaaS playbook.

ENTERPRISE/TECH NEWS

Mythos is here. Anthropic launched Claude Fable 5, its most capable generally available model, alongside Claude Mythos 5 for a restricted group of vetted cybersecurity and biology researchers. Fable 5 tops nearly every major benchmark and, per Stripe, compressed months of engineering work into days. The more interesting story is the safety architecture: Anthropic is explicitly tiering access based on risk, using classifier-based fallbacks rather than hard refusals, and requiring 30-day data retention on Mythos-class traffic. For founders building on top of frontier models, the access tiers and usage credit structure are worth reading carefully.

Datacenters in space. SpaceX disclosed in its IPO filing that it sees itself as the only company with a commercially viable path to orbital AI computing at scale, and is now reportedly targeting initial demonstration missions by late 2027, ahead of the 2028 timeline in its prospectus. The company has also filed for regulatory approval to deploy up to one million space-based data center satellites. The catch: the whole roadmap depends on Starship achieving rapid reusability at scale, which it has not yet demonstrated. For investors, this is a reminder that the orbital compute story is still largely a bet on Starship execution.

Legal AI fight. From Sacra: Legora, a Stockholm-based legal AI company built on top of frontier models rather than proprietary fine-tuned ones, hit $100M ARR in April 2026, up from roughly $50M at year-end 2025, and is now valued at $5.6B. It won its first major clients in Europe on data residency grounds before pushing into the US, where it now competes directly with Harvey ($300M ARR, $11B valuation). The more telling dynamic is what's happening at the firm level: rather than picking a single vendor, law firms are contracting with both Harvey and Legora and rotating seats between lawyers depending on the task. That hot-swap behavior suggests the legal AI market is fragmenting by workflow and practice area, not consolidating around a single winner.

HOW TO STARTUP

Is this AGI? Ethan Mollick put Claude Fable 5 through its paces on complex, multi-hour tasks and came away convinced it represents a genuine step change, not an incremental improvement. He built a sophisticated isochrone travel map, had the model spawn dozens of research agents to gather real data, and used it to produce a research calibration tool that took nine and a half hours to build. His broader observation is the more interesting one: the model has shifted the human role from operator to patron. You describe what you want, pay for it, and judge the output. The process itself is a black box, and the more capable the model, the more that may simply be the deal.

The end of cooperation. Noted on X that Anthropic will limit Claude Fable’s effectiveness for requests targeting frontier LLM development (e.g. building pretraining pipelines, distributed training infrastructure or ML accelerator design). And these degradations will not be made visible to the user so they won’t know. This is a phase transition in the relationship between the frontier labs and each other, and the frontier labs and the broader research community and their user base.

China has a robotics winner. Excellent analysis from SemiAnalysis: Unitree is approaching its IPO having shipped roughly 10,000 humanoid robots, with its G1 now crossing the threshold of economic viability for real warehouse deployments, undercutting human labor costs even under conservative, fully-teleoperated assumptions. SemiAnalysis argues the company is running the same playbook as BYD and DJI: own the critical component (in this case the actuator), bootstrap on hobbyists and researchers, then use that volume to fund iteration until the product unlocks commercial markets. For founders and investors watching the robotics space, the key number is $27,300: the current G1 list price, at an estimated 67% gross margin, still falling.

HOW TO VENTURE

Venture has to evolve. David Beisel, of NextView, has launched a public blog to run what he calls the diligence memo the industry refuses to do on itself. His argument is that most VC firms are treating AI as a sustaining innovation: faster sourcing, better analysis, same fundamental craft, which is exactly the move an incumbent makes before getting disrupted. He thinks AI is coming for the work itself, not just the tools around it, and that the reflex defense ("venture is about relationships and conviction, not information") sounds uncomfortably close to what every incumbent says on the way down. Worth reading for any investor who has made that argument recently.

A monthly VC market dispatch from Pavel Prata surfaces several data points worth sitting with. Roughly 85% of 2017-2018 vintage funds still haven't returned 1x DPI after seven or eight years, with median distributions at $0.34 on the dollar despite respectable paper IRRs. The SpaceX IPO could be the single largest DPI event in venture history landing into that drought, with Founders Fund's early $20M check potentially returning over $60B. Meanwhile, the largest funds have more than doubled their early-stage deal activity in the AI era, crowding into top-decile rounds and leaving average seed territory to emerging managers, of whom 78 new US funds launched in Q1 2026, up 34% year over year.

Moonshot features on TheMarker’s list of innovative startups of the year.

CruxOCM
Pipeline Hydraulics Engineer (Vancouver, Remote)

FalkorDB
Core Developer (Tel Aviv, Hybrid)

Reco
AI Software Engineer (Tel Aviv)