Newsletter.exe - June 2025
A monthly inside look at emerging technologies, innovation funding, and the real work of early-stage venture capital.
Heads up: this newsletter is deliberately long-form and may be truncated in your inbox. To read the full version, click "View entire message" or open it directly in your browser by clicking the title above. This letter was originally written for our LPs, we’ve decided to make it public. If you enjoy it, feel free to share it with friends or colleagues.
Dear all,
A quick summary of what you’ll find below:
Trends & Explorations
Investment & Portfolio
Community
Field notes
Current reflections
1. Trends & Explorations
Signals from the frontier: what we’re tracking, and thinking about
This section surfaces early signals that shape our curiosity and, sometimes, our theses. It’s where we map shifts in infrastructure, interface, and intent, before they solidify into obvious categories. These aren’t polished predictions, but rather working hypotheses grounded in what we’re seeing in decks, research, and founder conversations.
Over these past few weeks we’ve had three things that stood out and led us to some reflections on where things are going:
Observability as Leverage in LLMOps
In the early days of MLOps, observability was mostly a safety net. In LLMOps, it’s turning into a competitive edge. As language agents permeate production workflows, being able to diagnose behavior, quantify hallucinations, and track context fidelity in real-time is now essential. The most interesting approaches don’t just log or visualize, they build feedback loops that let teams fine-tune prompts, adjust context strategies, or switch models dynamically.
Agentic Product Management
Product management is undergoing its own automation wave. The latest agentic-native PM tools treat Product Requirements Documents (PRDs) and backlogs as living documents, co-written by humans and software agents. They don’t just organize tickets, they read codebases, monitor execution, suggest scopes, and generate spec drafts. As dev agents mature, a new orchestration layer should emerge: one where coordination, rather than task execution, becomes the high-leverage function.
MCP (Model Context Protocols): The New Strategic Middle Layer
Beyond traditional APIs, a new interface standard is emerging: Model Context Protocols - if you are a regular reader, this won’t feel new. MCP lets developers serve dynamic, structured, and persistent context to foundation models across use cases and applications. Sitting between app logic and inference layers, they abstract things like memory, grounding, and user state. It’s a powerful shift: apps can now treat reasoning as a service (modular, configurable, and external) rather than something built from scratch each time.
Rebuilding Payroll as Core Infrastructure
Most vertical SaaS tools still layer sleek interfaces over legacy workflows. The new wave is different: they’re rebuilding the stack around core value, not just features. Take payroll, for example. These next-gen engines don’t stop at UI: they offer transaction-level control, model-based reconciliation, and system-wide transparency. What makes this moment unique is a rare alignment of incentives: firms want more control, clients want clarity, and legacy vendors (through price hikes and flat roadmaps, caught in a typical Red Queen race) are actively pushing users away.
Oral Biologics and Localised Therapeutics
A paradigm shift is underway in gut-targeted therapies: oral biologics are no longer constrained by degradation or delivery inefficiencies. New encapsulation tech now allows biologics to be delivered directly in the gastrointestinal tract, preserving function while reducing systemic exposure. The implications are vast, from Irritable bowel syndrome (IBS) and Inflammatory Bowel Disease (IBD), to microbiome modulation and food allergy treatment.
Revenue Automation as Systemic Infrastructure
The next frontier in revenue tech isn’t prettier dashboards, it’s agents within your order-to-cash infrastructure that reconcile quotes with POs, chase payments, triage edge cases, and learn from escalation patterns. The real moat is below the interface: persistent memory, adaptive workflows, and LLMs trained on customer-specific behavior. As financial operations become more agentic, the winners may not look like SaaS, they’ll look like embedded infrastructure, invisible but indispensable.
Team.exe
2. Investment & Portfolio
From thesis to traction: what we’ve backed, and how it’s unfolding
A quick overview of the latest investments we’ve closed or committed to, and highlights from across the portfolio. We use this section to trace how our conviction translates into action, and how the companies we support are pushing their edges, whether through scientific validation, product velocity, or market resonance.
New Investments
No new closings this month (the two most recent accepted offers are still in progress).
We made three new offers that have been accepted in June: one in plant biology, one in AI for product management, and one in AI video game tooling.
Portfolio
Status: 29 companies
Highlights:
Exxa hosted the first vLLM & inference meet-up in Paris to accelerate the GenAI infra ecosystem: 70+ engineers joined for deep dives with AMD, Scaleway, and more. Details in this post. Details in this post.
Moyai repositioned as the go-to analytical AI agent for data-driven XOps teams, embedding fine-tuned reasoning models directly into platforms like Databricks to eliminate hallucinations and unlock native-format insights at scale.
Jimini partnered with the Paris Bar to offer thousands of small law firms (1-20 lawyers) 3 months of free access to its generative AI platform, designed for French and European legal systems. Details in this post.
3. Community
Our operating system: where the conversations happen
Our lunches, dinners, and shared spaces are more than just gatherings, they’re where hypotheses get sharpened, patterns emerge, and operators meet on equal footing. This section gives a glimpse into what’s being debated, discovered, or co-invented across the Galion.exe community. And if you, or someone on your team, wants in, just ping us.
🇳🇱 Amsterdam.exe {2}: we’ve gathered again in Amsterdam with a great group of European and local tech builders. This time, co-hosted a Summer Party with Tech Makers, who are preparing something really cool, and ran a hands-on sales workshop led by Guillaume Duvaux. Good energy, sharp minds, and lots of follow-ups.
TechLunches.exe with Charles Bochet (Twenty), Daniel Jarjoura, Daniel Madalitso Phiri (Weaviate), Noé Achache (Theodo), Duong Nguyen (Ekimetrics), Olivier de la Clergerie (Groupe LDLC), and Vinayak Mehta (June.so).
Topics were :
Defining AI Agents vs Workflows
Voice Interfaces and Natural Interactions
Forecasting and Reflexivity
Open Source Model
🥗MarketingLunch.exe with Adil Mania (Argil), Alex Kudelka (Outdo), Anastasia Wolter (EF), Caroline Roullet (Vivatech), Diane Bruneau (The Galion Project), Kevin Straszburger (Dust), Maxim Poulsen (Contrast)
Topics were:
Automation, Agents, and AI Use Cases
Content, Authenticity, and Format Innovation
Cognition, Learning, and the Future of Work
Channels, Saturation, and Growth Tactics
Bonus: most used Dust agents
comprehensive summary here (including tools)
🥗SalesLunch.exe with Gaëtan Gachet (Algolia), Hugues Decosse (Qantum), Charles Dusser (Alan), Hugo Benloulou (Synaps, xManoMano), Raphaël Arroche (Jimini), Tiger Solomons-Tibi (stealth), Julia Denoly (EF).
Topics were:
managing S-curves to keep steady growth
scaling SMEs channel
indirect vs direct sales force
incentive plan & the benefits of team bonus
hiring sales
🖥️ SOTA: we’re relaunching SOTA this fall with new formats and a new partner. More details in September. Very excited for what’s coming!
🧬 Frontier.exe: we’re kicking off Frontier.exe, a new format for deeptech founders to exchange openly on challenges, experiments, and emerging insights. No panels, just real conversations among top builders.
4. Fields Note
What we’re reading, underlining, and passing around
This is our living archive of influence: articles, papers, and essays that left a mark. Some sharpen our theses, others challenge our assumptions or offer language for half-formed ideas. We share them here not as endorsements, but as resources we think are worth your time.
Artificial Intelligence
Training Compute of Frontier AI Models Grows by 4-5x per Year (Epoch AI): Epoch’s data shows that the raw compute used to train leading AI models has been scaling at ~4.5x annually since around 2010. Meanwhile, algorithmic advances are making that compute more efficient (approximately 3x efficiency gain per year), meaning effective compute, which combines raw power and smarter methods, is rising at over 10x per year.
The cost of AI is decreasing (Ramp): businesses were paying $10 for $2.50 per million tokens in a year, a 75% drop. Improved AI efficiency and fierce competition among providers (OpenAI, Anthropic, etc.) drive this decline. Though energy remains a major cost, gains in model/infrastructure efficiency suggest token prices may keep falling, possibly making AI as cheap as a Google search
Frontier Lab Performance Converges: The era of model supremacy may be ending: the GPT-4 barrier has been significantly surpassed by 18+ models from various organizations. Google, once seen as lagging, now has 2 Gemini variants in the top 3 spots. OpenAI and Anthropic remain strong, while new entrants like DeepSeek, xAI (Grok), and Alibaba (Gwen) are climbing. Mistral is slipping in benchmarks despite earlier momentum.
How 100 Enterprise CIOs Are Building and Buying Gen AI in 2025 (a16z): a16z surveyed 100 CIOs across 15 industries and found Gen AI moving from pilot funds into core IT and business-unit budgets, with spend projected to grow ~75% year-over-year. Enterprises are deploying multiple models (OpenAI, Google, Anthropic dominate), often using 5+ models in production to optimize cost and performance. Procurement processes now resemble traditional software buying: rigorous evaluation, hosting scrutiny, benchmarking, and increasing switching costs. Off-the-shelf, vertical AI apps are overtaking custom builds, as firms prioritize fast, third-party solutions over in-house development
AI Signals the Death of the Author (Noema, David Gunkel): philosopher David Gunkel argues that LLMs challenge the traditional notion of authorship. Since AI-generated text has no singular origin, meaning shifts from creator to reader. In this new paradigm, interpretation replaces intention, the author fades, and the audience takes center stage.
Venture Capital
On the power law of Y Combinator startups (Jared Heyman): analyzing YC’s portfolio reveals a classic power law: a handful of breakout startups drive nearly all the returns. Of over 3,000 companies funded, just 1% represent 70% of the total portfolio value. This concentration has deep implications: YC’s success is less about average quality and more about consistently spotting a few outliers. For investors, founders, and accelerators alike, the takeaway is clear, success often hinges on being in the room when lightning strikes.
The Gift and The Curse of Staying Private with Bill Gurley (Patrick OShaughnessy, Invest Like The Best - Youtube podcast): legendary VC Bill Gurley outlines a system-level reset in venture: mega-funds dominate, 1,000+ “zombie unicorns” linger with outdated valuations, IPO/M&A windows are shut, and LPs are liquidity-starved. AI is both platform shift and distraction, fueling speculative funding cycles eerily similar to the ZIRP era. Gurley warns: without a correction, the ecosystem favors burn-heavy blitzscaling over sustainable company building.
The New Math: Why Seed Investors Are Selling Earlier (TechCrunch): (some) seed funds like Precursor Ventures are offloading winners as early as Series B to return capital faster. With LPs wary of 7-8 year holds and secondaries more accessible, smaller VCs are trading home runs for 3x outcomes and liquidity. Less patience, more engineered exits. Smart play or short-termism? Time will tell.
The Contrarian Blueprint of Founders Fund (The Generalist, Mario Gabriele): Since its inception in 2005 by Peter Thiel, Founders Fund has defied industry orthodoxy: no sector mandates, no ownership targets, and a deep skepticism of consensus thinking. From early bets on Facebook, SpaceX, and Palantir to recent contrarian calls (crypto during winter, defense tech amid skepticism), the firm has built a $12B+ portfolio around bold, nonconformist convictions. Its structure prioritizes speed, alignment with founders, and long-term capital, eschewing conventional fund cycles. But its libertarian ethos and Thiel’s polarizing presence raise questions about internal dynamics and external influence.
Innovation
A Creativity Lesson from Betty Crocker (Psychology Today): a failed instant cake mix becomes a creativity case study. When consumers rejected the product for being too easy, marketers removed the powdered egg and asked users to add a fresh one — restoring a sense of effort and ownership. The lesson: people value creation more when they feel personally involved. It’s the “egg theory“.
Agency Is Eating the World (Gian Segato): this essay argues that the real divide in the AI era isn’t technical skill but agency, the mindset to act without permission. With AI lowering barriers to execution, generalists and solo founders can now outcompete specialists. Startups like Midjourney and Super.com prove that tiny teams (or even individuals) can build at scale. The new edge? Not credentials, but the drive to just start.
The Race to Become the System of Action (Tidemark Cap): David Yuan explores how vertical SaaS platforms are shifting from passive systems of record into proactive systems of action, tools that both initiate workflows and execute tasks. He frames the evolution around four concepts: “control points” (hero users with high engagement), engagement as a strategic lever, the dual roles of running versus doing the work, and the emergence of systems that don’t just record but act. Success hinges on embedding into high-frequency workflows and offering agentic capabilities that “do the work,” not just facilitate it. Yuan predicts a race in vertical SaaS: platforms that evolve into true systems of action will own workflows, drive user behavior, and dominate their industries.
Generative AI Reduces Engagement and Memory in Student Writing (MIT Media Lab, arXiv): this study shows students using ChatGPT to write essays had lower cognitive engagement and struggled to recall or explain their own texts. Generative AI lightens the load, but erodes memory and authorship. Researchers suggest adding “intentional friction” (like prompt justification or recall quizzes) to rebuild agency and make creation more active.
5. On our mind.exe
What’s emerging when we zoom out
This section is where we take a step back. Sometimes it’s a reaction to a big ecosystem signal, sometimes it’s a sketch of where our thinking is headed. We use it to surface blind spots, tension points, and early frameworks, often imperfect, always exploratory. Think of it as the scratchpad behind our investment thesis.
Opening up the scratchpad
This month, we’re doing something slightly different. Instead of sharing a line of thought, we’re opening up a section of the deck we presented to our LP Advisory Committee (think: the board of directors, for a fund).
We believe ideas get sharper when shared early, and in good faith. By publishing this extract, we hope to contribute to a more open and grounded conversation about what’s really happening in the AI and VC ecosystem, beyond the hype cycles and behind closed doors.
The first part of the extract focuses on AI: the performance plateau of frontier models; the price-collapse flywheel powering mass inference; the strategic edge of being the biggest and having distribution; and the increasing difficulty of staying first. Most categories now see rapid turnover in performance leadership, across vision, video, and language, with voice as a notable exception, where user familiarity and switching costs provide relative stability. We also highlight the shift from model supremacy to orchestration and agent frameworks; the growing fragmentation of the stack; and the changing priorities of enterprise buyers, from accuracy to cost, security, and compliance.
The second part shifts from technology to the investment environment, with a focus on venture capital. We look at the structural polarization of the market, where fewer funds are being raised, but with larger sizes and more concentrated capital. We map the growing backlog of zombie unicorns and the fragmentation of liquidity pathways (from continuation funds to invite-only secondaries). This is creating pressure on early-stage timelines and Series A conversion, not just because capital is more concentrated, but also because the narrative has moved on. Many pre-A companies from previous cycles now face a harder fundraising environment in a market increasingly focused on GenAI-native plays.
We’ve spent some time pulling these threads together and figured they might be useful to others thinking through the same shifts. If you have contrarian or complementary views, we’d love to hear them!
Stay curious, keep learning,
Willy