Brian Schechter

The through-line

Across everything Brian Schechter says runs a single test he applies to founders: can you see the whole board? He returns again and again to the idea that what separates a memorable AI infrastructure pitch from the dozens that blur together is "someone who can really clearly articulate a vision and strategy and like see a chessboard who like who the players are how the Market's going to evolve and have a strong opinion on what's going to unfold and why the things that they're building are going to take advantage of how those things are going to unfold" [1]. Underneath that sits a set of first principles he treats as non-negotiable: "what's the problem you're really trying to solve here who's the customer why are you uniquely well positioned to do that and if you don't have great answers to those things you're going to run into walls" [1].

The second, related conviction is that this kind of clarity is only useful if the investor on the other side of the table has done the work to evaluate it. He argues that "investing is incredibly competitive right now and you no longer can be a generalist and expect to really understand what is best in class" [2], and that a firm's real contribution is not advice but labour: recruiting and revenue [2]. His attention has moved with the market, from consumer and social platform businesses [5][10] through the economics of running data on Kubernetes [8] to the current preoccupation with compute, models and the infrastructure layer between models and applications [2][3].

On what makes a pitch stand out

Schechter's diagnosis of bad pitches is not that founders are inarticulate, it is that clarity pushes everyone toward the same words. A founder "need[s] to like speak a language that VCS are going to understand," and "the problem is that what that ends up leading to is you explaining your business the same way like a bunch of other Founders are explaining it" [1]. The balance he wants is being "very clear very sort of like direct and genuine about what it is you're building but not lose sort of a relatively untechnical unsophisticated investor in like the details of like what you're what makes it different" [1]. He concedes this is unusually hard in AI infrastructure, "because like everyone's doing things that seem kind of similar and that like you know in 5 years what you're doing now could bleed into like what some other company is doing and like the lines are really blurry and the stack is still kind of like evolving and being formed" [1]. In practice, he says, the recognition is blunt: "it's like the number of times we at a pitch it's like oh wait a second this is sounding a lot like yeah the pitch we heard earlier today and yesterday and last week" [1].

What he is actually screening for is what he calls, with acknowledged cliché, a learning machine, and he is honest that this is "less a advice for Founders and more reflection for us because if that's not you as a Founder it's very hard to fake" [1]. The chessboard answer works because it demonstrates that "someone is engaged in a process of deep thinking about their business and you can tell that it's going to just keep on changing" [1]. Investors would prefer to invest in lines rather than dots, he notes, but "oftentimes you can't you don't have that luxury" [1], so the strategic narrative substitutes for the missing time series. The controllable part is authenticity: he advises founders to be "less perfectly Polished" and instead "let me bring you into my thinking let me tell you like I had this insight and then I had this insight" [1], landing on deeply held beliefs and a strategy while making clear the game plan will keep adapting, "you don't know the whole chess game at the beginning you have to keep adapting" [1]. He also frames his own public writing as an antidote to the opacity of the process, since "as a Founder understanding like why an investor thinks certain things are interesting versus not interesting is very opaque and seemingly kind of like irrational and idiosyncratic" [1].

On tailwinds and the anatomy of acceleration

He is unusually direct that conventional projections are the least interesting page in a deck. What matters more than saying "in 12 months or 18 months we're going to be at a million in AR or 2 million AR and that's because our acvs are X and we have this number of customers" is "demonstrating a like firm grasp of what it would mean for you to accelerate far beyond that from like a what are the Tailwinds perspective" [1]. His evidence is a set of trajectories that look impossible on paper: MosaicML going from zero to $30 million in roughly six months, weights and biases going from about 2 to 40 over eighteen months, and together reaching north of 20 in under a year [1]. He accepts that "no one puts that on their on their slide deck on year one 0 to 30" because "you'll look crazy," but insists founders "do want to be able to understand why your business is so exciting that it has the possibility to accelerate like that" [1].

The lesson he draws is about conditions rather than effort. Those stories happen not "cuz it's just that someone built something great," but because "someone got the right the timing right" [1], and the questions he pushes founders to answer are what tailwind they can ride, how they position their tech and brand, who the ICP is, and "why are they going to be so looking to buy so quickly" [1]. He pairs this with permission to be patient: MosaicML itself spent roughly eighteen months in a research phase serving researchers and academics before going commercial, so "you need to know where you are" [1], and if hypergrowth is not your situation, "you might need to be building for a long time" [1]. Finding the right investor, in his framing, "is just like finding the person who believes that there's actually like a chance of that story coming to fruition" [1].

On the Databricks acquisition of MosaicML and the LLMOps stack

He treats the $1.3 billion purchase of MosaicML on roughly $30 million of ARR, for a two-year-old company that had only begun serious monetisation months earlier, as a staggering but legible transaction [1]. MosaicML's founding insight, as he heard it from chief scientist Jonathan, was that training a foundation model leaves "all this like wreckage" of failed experiments whose compute and human cost is worthless once the lesson is extracted, so capturing those lessons and selling them onward would amount to "democratizing uh generative AI" [1]. That matters because the demand he sees from large and increasingly mid-market companies is not for API-based closed models but for "taking open source models and fine-tuning them with their own internal data," which "enables a degree of flexibility and control that's really valued by larger companies" [1]. MosaicML was the leading route to fine-tuning an open source model and running it on your own infrastructure, and the missing piece in Databricks' end-to-end LLMOps offering was exactly that ability to train and refine your own model [1].

The complication he flags is customer fit rather than product fit. Databricks has long sold to large corporations "with really big engineering teams maybe not the most like Tech forward businesses but big companies with big teams and lots of budget," while MosaicML's revenue came from "emergent fast growing tech companies that wanted to embed AI into their product offering," with Replit as the emblematic account [1]. His hunch is that the Venn diagram overlap is small, which means MosaicML will likely be pushed upmarket and absorbed into the Databricks product, leaving an open question about its original target customer and, by implication, an opening created for startups [1].

On repositioning fast enough

Databricks also serves him as a case study in narrative agility. He points out that the company's website at the end of 2022 and the same website today read like "two different companies," one about your data lake or lakehouse and the other about your AI [1]. He calls the mission consistent, enabling the enterprise to leverage the value of its data over a dozen-plus years, and finds the adaptation remarkable given the company was already north of a billion dollars before the LLM explosion [1]. The founder-facing lesson is that it "may seem obvious" but is "actually very important," because "it's very easy to get super committed to what you're doing and Miss massive Market changes um and adapt slowly to changing your positioning" [1]. He says he now watches, in portfolio founders and in pitches alike, "how fast is their speed of iteration around how they present what who they are to the world" [1].

On the crowded market and the AI gorillas

He describes the current moment as a gold rush, with "college students drop out of school to like go builds in and around geni" at a rate beyond anything he saw in a decade as an investor and founder, and with researchers who have spent decades on the technology now taking their swing, so "Talent is coming kind of coming from like all places" [1]. The result is saturation in one particular pitch: "I'm going to be the one to enable this company to Leverage The Power of AI," which he says he sees "very very crowded" [1]. His response is bracing. "A piece of me almost wants to say like if you're doing that do something else," because a set of "AI gorillas that are going to control so much of the AI Market" are already off to the races from 2018 through 2023, and "there's already a lot of sort of crud that is emerging where businesses are derivative of what emerging leaders look like" [1]. Chasing the wave behind weights and biases or together is "already a hard wave to be going after without something without a really compelling reason to win" [1].

Where he does see room is at the infrastructure layer rather than in the thesis of rebuilding SaaS applications with generative AI and undercutting incumbents on price, which he says he understands but is less interested in [1]. The question that engages him is "who are going to be the people to really rebuild existing um sort of table Stak Solutions zapier data dog," and whether the generative AI version of Datadog turns out to be Datadog, an outcome he says depends entirely on "that business's strategic vision and execution" [1]. He frames the whole moment as risk and opportunity arriving together: a powerful technology whose best use is still unclear, enterprises exposed, and startups positioned to take advantage of incumbents that fail to embrace it [1].

On where value accrues across the AI stack

He refuses to pick a layer. "The value for AI will occur throughout the entire stack. Compute absolutely models for sure. Infrastructure between the models and applications yes and the applications themselves" [2]. He specifically rejects the argument that the model layer is closed, noting that people once said it was settled among a few winners and that "that's clearly no longer the case especially with Deep Seek," alongside founders building their own distinct foundation models, with new players likely at the compute layer too [2]. He also expects AI's reach to extend past knowledge work into construction and military applications, and argues that delivering those solutions will itself require infrastructure [2]. On the application side, he sees "every single industry" getting a vertical AI solution that goes beyond helping people do their jobs to "actually providing automation to pro deliver services and complete jobs themselves" [2]. His answer to the fear of a saturated problem space is structural: as AI automates today's problems, "it will unleash sort of new problems to solve," because technology's limits keep getting pushed [2].

On compute and hardware as the advantage

Compute is, in his words, "perhaps the hottest area of startup land right now," and he frames the present as "an interesting moment in history where all of a sudden the the hardware matters" [3]. The conversation he builds around that premise concerns the end of computational assets as commodities, the need for tight integration between hardware and software, the waste in current server and power supply design, and the argument that order-of-magnitude gains in speed, cost and efficiency are unreachable without hardware innovation, alongside a broader claim that diversity in both hardware and software drives innovation [3]. The economics beneath infrastructure adoption is a recurring interest of his more generally, including the business and economic drivers that accelerate or block running stateful workloads on Kubernetes, and how that space evolves alongside other open source trends [8].

On what a venture firm should actually do

His view of firm design follows from his view of the market. Specialisation comes first: Primary has people "focusing exclusively of the infrastructure layer, at the application layer, for the enterprise, for SMB, for healthcare, for fintech," with AI considered "first and foremost when we're looking at a deal" [2]. The point of that depth is that the firm only pursues deals where "we understand the the pain point of the customer" and the competitive landscape "with a lot of granularity," which is what makes it possible to help a founder navigate what is coming [2]. Primary itself he describes as an early stage New York firm with a billion dollars under management that has been building out its infrastructure practice over recent years [1].

On support, he draws a sharp line against the industry norm of "advice maybe some connections." His formulation is that "our currency is actual work": embedded recruiters and business development representatives who build pipeline for portfolio companies, plus financial modelling to prepare for a Series A, concentrated in the window right after a seed round when a company is trying to accelerate toward an A [2]. Fundraising help is table stakes; "the real areas where we're like moving the needle is around recruiting and revenue generation," which he calls "the two most important things when you talk to a founder" [2]. He roots this in his own experience as a founder Primary backed, where he had "incredible investors" on the cap table but Primary "was the only investor that felt like a team member" [2], an "uncharacteristic commitment to working alongside founders" [2]. He extends the same standard to the firm's own backers, citing Vintage's value-plus programme and its introductions to corporate leaders as an LP that genuinely delivers [2].

On earlier bets: platforms and personal branding

Before infrastructure, his attention was on the consumer side of the same structural question: what happens when the value you build sits on someone else's platform. At SelfMade he was working on tools and services for small businesses on social media as commerce moved from websites to Instagram accounts, and he has discussed both where social media was heading and the particular difficulty of building on a closed and constantly evolving platform [5]. The animating interest he brought to that work was in watching people grow and supporting their own version of self-expression, which he framed as a personal branding revolution [10].

Takeaways

• The pitch quality he rewards is strategic, not financial: a founder who can name the players, predict how the market evolves, and explain why their product benefits from that evolution stands out because "most Founders don't think that way" [1].
• Answer the first principles or nothing else helps: problem, customer, and why you are uniquely positioned to win, otherwise "it doesn't matter how good the story gets or how hard you try you're going to run into walls" [1].
• Extraordinary growth comes from conditions, not effort; study the tailwind, the ICP and the timing that would let you go from 0 to 30 like MosaicML, 2 to 40 like weights and biases, or past 20 in a year like together [1].
• Speed of repositioning is a diligence signal: Databricks' website changed so completely between late 2022 and today that it reads like a different company, and he watches founders for the same agility [1].
• The enterprise appetite is for fine-tuning open source models on internal data for flexibility and control, which is why MosaicML closed the missing training and refinement stage in Databricks' end-to-end LLMOps offering [1].
• Beware derivative positioning: the "I'll bring AI to the enterprise" pitch is saturated, AI gorillas are already off to the races, and chasing them without a compelling reason to win is a bad trade [1].
• Value will accrue across the entire stack, compute, models, middle infrastructure and applications, and the model layer is not closed, "especially with Deep Seek" [2].
• Generalist investing no longer works; deep sector and stage specialisation is what lets a firm judge what is best in class and help founders navigate the landscape [2].
• A firm's contribution should be measured in work, not advice: embedded recruiters and BDRs, pipeline, and Series A financial modelling, because recruiting and revenue are what shorten the path to the next round [2].