About
Tech founder in Paris, SF and NYC (4 exits) turned early-stage VC. Now investing Serena's…
Articles by Guillaume
Activity
10K followers
-
Guillaume Decugis shared thisIs Nvidia’s CUDA moat finally eroding? While Nvidia is a chip company, a big part of its massive competitive advantage has been software: the ecosystem, not just the silicon, that drives most AI to be written for their chips. Anyone who can loosen that grip stands to capture a lot of value... And a few recent signals are interesting in that regard: ➡️ Google, with Meta, open-sourced TorchTPU, a native PyTorch backend for Google’s TPUs, their own AI chips. The goal: take code written for Nvidia GPUs and run it on TPUs with a one-line change. ➡️ AWS released TorchNeuron in public beta, so PyTorch code can run on Trainium, Amazon’s own AI chips. Same idea. ➡️ Tenstorrent raised $1.5B at a $6.2B valuation to build chips for serving open-weight models. Native PyTorch backends on other chips lower the cost of trying an alternative to Nvidia’s GPUs. And some investors clearly believe other chip makers like Tenstorrent can make it. It’s still early, as the builders themselves say. TorchNeuron has gaps, and TorchTPU says peak performance still takes hardware-specific kernels. Getting a model to run is getting easier. Getting it to run fast is the harder part. That’s the layer we’ve been drawn to as VCs focused on foundational tech: compilation, or translating a model into instructions a specific chip can run efficiently. It’s hard. I still remember from my days in the mobile industry how painful it was to rewrite and recompile software from chip A to chip B to reduce our dependency on a single chip vendor. But to decouple software from hardware and grab a piece of the immense value Nvidia captures, someone has to map each model onto each piece of hardware efficiently. On our end, that's the rationale that got us to back Roofline, a foundational tech startup founded by Jan Moritz Joseph, Dr. Thomas Zimmermann, Maximilian Bartel and Rainer Leupers which tackles this problem first for edge devices. There's obviously a lot that needs to happen before the industry's dependency to Nvidia weakens but I'm curious: how long until a team picks its AI chip on price and availability instead of ‘whatever runs our code’? One year, five, never?
-
Guillaume Decugis shared thisIt almost feels like the longest part in Emmi AI’s amazingly fast journey was to plan the closing dinner… ;-). But we finally had it yesterday in Vienna! 🎉 It was great to pause a little and celebrate the unbelievable achievements of Dennis, Johannes, Miks and their team. Capitalizing on years of research to radically transform engineering with AI. 🙏 Incredibly proud and grateful to have been a part of this. 🔥 Starting from co-leading their seed round early 2025 as Juliette and I were looking for foundational models that leveraged physics - an assumption in our investment thesis together with Bertrand, Floriane and Matthieu being that there were interesting value to create away from the big labs in non-language modalities - as we also saw with pyannoteAI and Waiv, formerly Owkin Dx. We had seen a couple of opportunities so we had a bit of a headstart but a trip to Vienna and some intense discussions with the team convinced us this was the a team like none others for that space (and yes there was a lot of whiteboarding to dig into how a field-based approach differed from a mesh-based one…). 🚀 Fast forward to this spring and the synergies were Mistral were obvious. Suffice to look at the amazing industrial clients such as Airbus or BMW Group that they announced publicly: a core unique foundational layer accelerated by high-tailwind go to market will transform the space at GPU speed. 👍 Grateful to our co-investors and board members Peter, Andreas and Laurenz for navigating this fast-paced, innovative journey with openness, intelligence and ambition. And thanks Thomas and Lukas for your deal support!
-
Guillaume Decugis shared thisWill decision models last as a new category? Right now, they're betting on a standard API to make it happen. In a few weeks, decision models went from one startup's experiment to a category, with OpenAI, Microsoft , Amazon Web Services (AWS) , Perplexity and Cloudflare all shipping their own. One schema is winning: many of them already speak like Jef, using TypeSafe AI's System One API (a choice, a score or a yes/no probability, no prose). It's the Chat Completion API playbook all over again. ➡️ For users, it’s great news. Swap Jev for Strands Decider or Kev by editing a URL. ➡️ For the first mover, debatable… TypeSafe now has to win on model quality alone. ➡️ For incumbents, likely resistance: OpenAI's Decisions API runs on the same GPT-6 Luna it already sells. Pushing to deny this being a new model class, but rather a mode of the LLM? 💰The case for "category" is the economics. A decision costs a fraction of an LLM call and takes a fraction of the time (OpenAI claims about 10x faster). Nobody wants to pay LLM prices for the hundreds of small judgments an agent makes. 👥 The case for "feature" is how fast this gets copied. Within about two weeks of Jev going public, Jared Palmer, Cognition's VP of Engineering, shipped Kev 1.0: open weights, the same API, built with a coding agent and an outside write-up of Jev's architecture. The core training run for the small models is about an hour on one H100. The biggest one is within a point of Jev on new-source questions, on his own benchmark. 🧬 Because at the core, most of these “models” are sitting on top of someone else’s. When most of the field is someone else's open base model plus a small decision head, is that a new category of model? Or a fine-tuning recipe with a standard API on top? As specialized VCs focused on foundational layers with defensible tech, we’ve found language a tough modality to build new models in, given the intensity of the competition from the big labs (we’ve had great success in non-language modality such as Physics with Emmi AI or voice with pyannoteAI)pyannoteAI). Whatever the outcome for TypeSafe, the above shows it won’t be an easy category to win, if it materially becomes one. To do so, I wonder whether solving the remaining unsolved questions would be enough: how reliably a model knows when it's wrong, plus evals and the data that sets your thresholds (A USC researcher flipped 61% of 508 correct Jev decisions with about 31 words of optimized context. That's an adversarial rate, not an everyday one. But still…). Curious to hear your thoughts on decision models? Will they last as a category? Do they provide enough surface for new foundational tech builders to leverage? A great recap on decision models by Janakiram MSV on The New Stack: https://lnkd.in/eTDQ7iJb Hugging Face's benchmark of decision models: https://lnkd.in/e3PTAuVk Cc Bertrand, Floriane, Juliette
-
Guillaume Decugis shared thisWhat will be to data science what agentic coding has been to software engineering? One answer just launched this week. Tabular data is where enterprises count, compare and decide: transactions, customers, claims, orders, all in rows and columns. Predictive models, built and trained by data scientists, turn that data into decisions, like which customer will churn, which claim looks fraudulent or how much stock to order. That makes the model the point where data becomes action, and it will stay so, because tabular data remains the ground truth behind every prediction. Of course, agents can now build a training data pipeline in seconds and a model in minutes. But a model that runs is not a model you can trust. A leaked feature or a flattering validation split produces a great score and a bad decision, and until now, nothing flagged it automatically. What catches those errors is judgment: knowing what to check, what to doubt and when a result is too good to be true. That is the craft of data science, and it's the scarce part. Agents speed up the typing, not the thinking... That's the bet behind Skore, the new enterprise platform by Probabl, the company behind scikit-learn that we backed last year. It lets data scientists build reliable tabular AI models faster with agents, with the guardrails and auditability that keep human judgment in control. Not agents instead of experts, but experts at agent speed. As agents multiply what enterprises build, the durable value sits in the infrastructure that makes their output reliable. That has been a core idea in our investment thesis, and Probabl is building exactly that. Kudos to François, Gael and the whole Probabl team on this new chapter. Learn more here: https://lnkd.in/eyCcmdUW Cc Bertrand, Floriane, Juliette, Matthieu
-
Guillaume Decugis shared thisFrom vibing to governing... What does it take to scale AI benefits and generate durable ROI in the enterprise? Every engineering team discovered the same magic over the past years: point a coding agent at a problem, watch it work. Now that we all witnessed that the thing works at all, we're now entering the industrial phase. Once agents are running in production across hundreds of engineers, "it works" stops being the question. Who's spending what, on which models, through which of a dozen harnesses that change weekly? That's no longer a problem of whether the magic works - it's an operations problem. And it needs a governance layer. Qonto just showed what that looks like in practice. 500 engineers migrated onto Edgee's Agent Gateway in six weeks: - Adoption went 10 → 350 organically before it became the default. - Combined compression and model routing cut their AI spend by 37.7% (without changing a single engineer's tools or workflow). In Kevin Prettre's (Qonto's Director of Infrastructure) words: "Edgee has become our governance layer for AI usage." The pattern is familiar: cloud, then FinOps; data, then data governance etc... Now: agentic coding, then its governance layer. ➡️ Sacha Morard and the Edgee team saw it coming a while back and now they're delivering real, tangible results. 🚀 Behind this early success story is what excites me most as an investor focused on foundational tech: the potential emergence of a new category. We're obsessed with finding the pieces of AI infrastructure the big labs won't build themselves - beyond wrappers, built for durable, long-term customer value. As agentic coding becomes the norm, we're excited to see Edgee play a key role in the governance layer that's emerging. Full interview of Kevin in the first comment. cc Bertrand, Floriane, Juliette, Pauline, Matthieu, Olivier, Jerome, Lily, Apostolos, Gilles
-
Guillaume Decugis shared this~93%. If you were wondering what AI labs are willing to pay for your data - which, in reverse, is also the cost of privacy - we now have a concrete data point. Meta just released pricing for Muse Spark 1.3, its latest state-of-the-art model, with two distinct tiers: 💸 Standard tier (zero data retention): $1.25/$4.25 per million tokens 🔗 Contributor tier (train on my data): $0.10/$0.20 per million tokens ➡️ That's a blended 93% discount if you let them train on your prompts. Tomasz Tunguz did the math across a 30:1 agentic input/output ratio (link below) and arrived at $1.24 per million tokens as the implied price of zero data retention (ZDR). For an enterprise running 1B tokens/day, that's a ~$454K/year privacy surcharge for staying private. Or, flipped around, a $454K/year data dividend for opting in. Before AI, business models split cleanly: 😀 Consumers got free products in exchange for their data ("you are the product") 👩💼 Businesses paid for SaaS and kept their data 😈 Meta is introducing a tempting twist in this model. Maybe it's consumer, data-savvy DNA. But mostly it's that the next leg of the frontier race runs through post-training and real usage patterns - and when specialized labeling platforms charge $75–200/hr for verified reasoning trace, trying to cut out the middleman is just good business. By turning its inference API into a data pipeline, Meta just built a self-funding alternative to Mercor and the likes. Compute stops being a utility you sell by the unit: it becomes currency you trade for training data. Given Meta's historical strength on consumer but lag on selling tech B2B (selling ads is different), this seems like an interesting bet - especially for SMBs who might not be that concerned about data sovereignty. One observation and two questions from me: 🧠 This is another sign of how much the infrastructure layer will shape the economics of AI and everything built on top (cc Bertrand, Floriane, Juliette), 💸 Will $1.24/m hold once OpenAI, Anthropic, or Google feel pressure to respond? Will they have to? 👩💻 Does this trickle down? If discount-for-data becomes the default expectation at the model layer, do startups building on top end up having to replicate the same barter with their own users just to stay price-competitive? Full piece from Tom Tunguz: https://lnkd.in/eDtvPte6
-
Guillaume Decugis shared this❤️ "I love it when a plan comes together." 🚀 Congrats Mistral, Arthur Mensch, Guillaume Lample, Timothee Lacroix for closing your €3Bn Series D at a post-money valuation north of €21Bn - nearly double where it stood a year ago, and the largest equity round ever raised by a private European tech company. 🥂 Near and dear news to me as we've seen two of our direct portfolio companies successfully acquired by Mistral: ➡️ Koyeb, the serverless AI inference platform. The type of elegant plumbing infrastructure frontier AI companies won't scale without - when you realize the model is half the solution without the specialized cloud to support it. ➡️ Emmi AI, Physical AI as the next frontier after language. Large Engineering Models compressing 48-72h simulations to seconds - an investment we made as we analyzed frontier labs would focus on language while there were plenty of other vertical modalities for specialized teams to build on. Same thesis of investing at the foundational layer. Two nice validations, now amplified by a +80% uptick in valuation in just 12 months since Series C. Infra and deep tech aren't the easiest part of AI to understand. But they're the foundational picks and shovels that keep compounding while everyone else is rushing to the application layer. And today's a good opportunity to remember how strong European founders are at building this foundational layer. So kudos as well to Yann Leger, Edouard Bonlieu, Bastien Chatelard, Dennis Just, Johannes Brandstetter and Miks Mikelsons. Together with Bertrand, Floriane, Matthieu and Juliette, we'll keep rooting for your ongoing success with the Mistral team! #AI #DeepTech
-
Guillaume Decugis shared thisHyperscaler as a service, anyone? $500B of third-party capital. Six independent financing platforms. Nvidia isn't just selling chips anymore - it's providing the financing and software core of a hyperscaler to neoclouds. Circular vendor financing, a bubble waiting to pop? Enough demand down the road? I'm sure skeptics will have lots of questions even though the capital is coming from BlackRock, Blackstone, Brookfield, KKR - independently underwritten, not Nvidia's balance sheet. In short, outside capital betting the demand is real enough to not need Nvidia to backstop it. But beyond that debate, I think it's interesting to look at what everyone else in the stack is doing at the same time: 🚀 Nvidia betting on HaaS, building the financing + software layer on top of GPUs to enable “hyperscaler in a box” (or part of it - securing land and power is still something few neoclouds do and hard-to-replicate expertise providing real value, as we’ve seen with metrobloks, our portfolio company) 💻 Big tech building captive silicon (TPU, Trainium, Maia) to defend their own 35-40% margins (Google TPU v7, AWS Trainium 3, Microsoft Maia 2 — all in volume production in 2026, ~1.9M accelerators combined) 🧠 Labs building owned infrastructure (OpenAI's Camellia, its own $20Bn data center, Anthropic's 10GW build) to stop being hostage to either ➡️ Everybody is trying to not be the one left dependent. Beyond wondering whether the demand shows up, it's fascinating to me to see that a new cloud topology is shaping up where it's no longer the big 3. Because it'll also come with completely new problems to solve and new foundational layer startups to fund. Btw a good read by Clark Tang on Nvidia’s latest move: https://lnkd.in/epqSA8d6 Cc Bertrand, Floriane, Matthieu, Juliette.
-
Guillaume Decugis shared this🚀 Excited to see Waiv, formerly Owkin Dx continue to push the frontier of foundation models for #digitalpathology in cancer treatment. AI is only as good as the data it gets trained on - which often can mean including a bias originating in data collection. Genomics with "batch effects" between sequencing labs, imaging (from radiology to astronomy) with sensor heterogeneity: models learning the fingerprint of the instrument instead of the signal it was built to capture. A lot of pathology foundation models similarly carry a hidden flaw: their representations often encode which scanner and which lab produced a slide, not just the biology in it. That's a real barrier to reliable, multi-site clinical deployment. 🧠 The Waiv team led by the amazing Meriem Sefta and Lionel Guillou just tackled this head-on with a model-agnostic fine-tuning recipe developed by Alexandre Filiot, Oskar Thaeter, Benoît Schmauch and Lionel Guillou that makes any pathology foundation model more robust to acquisition factors - without retraining from scratch. Applied uniformly across ten different foundation models, it improved every single one, lifting robustness by 23% on average and downstream performance by 43%, with no trade-off between the two. Better yet, the team is releasing two of the fine-tuned models to the community: Phaet (a robustified Phikon-v2) and Mascaret (a robustified Midnight-12k). A great example of rigorous, open science pushing AI-powered digital pathology closer to dependable real-world use. And for us as deeptech specialized investors, this is the kind of work we look for at the intersection of #AI and #lifesciences: not a thin application layer, but deep, defensible science that becomes infrastructure for the whole field to build on. Learn more with the article linked in the first comment. cc Bertrand, Floriane, Matthieu, Juliette
Guillaume Decugis replied to a comment
2d
Jan Moritz Joseph Good point. Open source can play a key role: the Linux story among others, right? ;-)
Guillaume Decugis commented on a post
2w
Carlos - Also: application layer (Hubspot, Monday…) vs infra / foundational layer (Snowflake, Datadog…). Low defensibility, easy to clone, easy to migrate vs high technology asset, high switching cost, platform value for AI, etc…
-
Guillaume Decugis liked thisGuillaume Decugis liked thisJ'ai lu l'amendement I-1358 au PLF 2027, adopté ce vendredi à l'Assemblée : une taxe de 0,055 € par Go envoyé vers les abonnés des FAI français. Le chiffre est abstrait, alors comparons-le. Corrigez-moi si vous voyez des erreurs. Deux comparaisons et un cas concret : 1. Le transit et le peering. Un port 100 Gbps coûte environ 6 300 € par mois chez un fournisseur de transit mondial. En interconnexion directe avec un FAI européen, compter 3 à 5 fois plus, soit 20 à 30 k€ par mois. Sur ce port (a), la taxe atteint 1,02 M€ par mois, soit environ 3400 % à 16300 % du prix d'une interconnexion vers les FAI français. On change donc profondément le modèle économique de la bande passante en France. 2. Les hyperscalers sont critiqués pour le prix de leur trafic sortant, qui va d'environ 0,08€ à 0,045€ par Go selon le volume. La taxe, 0,055 € par Go, représente 68 % à 123 % de ce prix. On parle donc de doubler, pour le trafic vers la France, un prix déjà jugé trop élevé ! 3. Le cas d'OVHcloud. Nous envoyons de l'ordre de 500 Gbps au 95e percentile vers les FAI français, soit environ 324 Gbps en moyenne. Sur un an, la taxe brute atteindrait 70 M€. Le texte la plafonne à 15 % du CA réalisé en France ou à 2 % du CA mondial, soit entre environ 23 et 70 M€ par an que nous devrions collecter pour l'État, et donc répercuter sur nos clients. Conclusions : 1. Une taxe porte en général sur une petite fraction de son assiette. Celle-ci vaut plus de 3000 % du coût du transit, et environ 100 % du prix public des hyperscalers. Rappelez-vous la loi SREN, votée pour supprimer les frais de transfert qui empêchaient les clients de quitter un fournisseur cloud. Ici, on parle de doubler le prix du trafic sortant vers les internautes français, un prix déjà très problématique. 2. Le garde-fou, un plafond calculé sur le chiffre d'affaires, est incohérent. Ce chiffre d'affaires couvre des serveurs, du stockage, des datacenters et des équipes sans rapport avec la bande passante. C'est comme créer une taxe calculée sur la consommation électrique des usines automobiles en France, plafonnée à 15 % du chiffre d'affaires des constructeurs. Quel rapport entre les deux ? 3. Une taxe nationale de ce type frappe à coup sûr les entreprises établies en France, alors que leurs concurrents mondiaux sans présence en France livrent le même trafic aux mêmes FAI français sans payer cette taxe en pratique. Cela crée une distorsion de concurrence au détriment des acteurs nationaux et déclasse les entreprises qui embauchent en France. Une taxe qui produit l'effet inverse serait une meilleure approche. (a) Facturé au 95e percentile, ce port n'est pas plein en permanence : avec un profil journalier type (22 % vers 4 h, 90 % vers 21 h), il transporte en moyenne 57 % de sa capacité, soit 18,6 millions de Go par mois.
-
Guillaume Decugis reacted on thisGuillaume Decugis reacted on thisAre you trading or have assets like batteries or are generating renewable power in Europe? EnergyBench, a new live benchmark released by Evgeny Genov and Prof. Christoph Bergmeir shows t0-gamma (with additional signals provided by the benchmark like weather) is #1 at forecasting day-ahead energy prices, load, and production for solar and wind in Europe power grids. Our model tops a very competitive podium with the recently released Timestral by our friends at Mistral and TimesFM3 from Google. We've seen consistent top performance on energy markets for our models since our first t0-alpha release, and we're excited to see live benchmarks like EnergyBench confirm this! Ioannis Boukas Bastien Fabre Kyle Gilson Sam Tegel Nicolas Ulmann Vincent Milville Reach out to us if you're using forecasts for prices, load and in general on the power grids - we can help! #forecasting #power #powergrids #energy #energytrading
-
Guillaume Decugis liked thisGuillaume Decugis liked thisWe are releasing the weights of a new and improved Mistral Large 4, advancing the open frontier in coding and cybersecurity. Trained from scratch with clear lineage in European data centers, Le Chonk provides frontier-level performance in enterprise-critical domains, and gives organizations a trusted foundation for Sovereign AI. 🛡Cybersecurity: Enterprise and public sector institutions can now download, deploy, and own frontier AI to detect, analyze, and respond to threats across code and IT infrastructure. The models can run on a company’s own infrastructure and no third party can restrict access or disable them. Industry-leading cybersecurity companies, including Cloudflare, CrowdStrike, Zscaler, Cisco, Elastic, Fortinet, Rubrik and Orange Cyberdefense, are supporting and testing ML4. 💻 Coding: ML4 delivers frontier coding performance with significantly fewer tokens. It leads Cognition’s FrontierCode 2.0, which measures whether a maintainer would merge the agent’s change, and matches the best open-weight models on the Artificial Analysis Coding Agent Index while using 47% fewer tokens per task. 💻 Work: ML4 also keeps improving on everyday enterprise tasks, with 73.3% on Finch for finance spreadsheet workflows and 84.5% on BrowseComp for web browsing. Mistral Large 4 will serve as the foundation for a new generation of specialized and open Mistral models. We’re committed to giving developers, enterprises, and institutions the power to own, customize, and control their AI.
-
Guillaume Decugis liked thisGuillaume Decugis liked this🚀 Internship Opportunity! The Content Seal team at Meta Superintelligence Labs (FAIR) is recruiting an intern to work on watermarking. Our research is deployed across Meta, and we have a strong focus on sharing our work with academia and open source, through papers, code, and models. 🔗 If you are interested, check out the job description and apply here: https://lnkd.in/gfKTFcc7 (or email me at pfz@meta.com with the subject line "Content Seal Internship Application") ℹ️ Details - Location: Paris - Role: Research Scientist Intern - Content Seal - FAIR - Qualifications: Currently has, or is in the process of obtaining, a Ph.D. degree in AI or a related field - Preferred experience: Experience in watermarking, forensics, security and/or generative modeling (text, image, audio or video) - Focus: watermarking, safety and security (fighting deepfakes and harmful content), training data provenance, etc. You can look here for an idea of what we've worked on in the past: https://lnkd.in/gi4biR5s
Experience & Education
Languages
-
English
Native or bilingual proficiency
-
French
Native or bilingual proficiency
Other similar profiles
-
Adam Singolda
Adam Singolda
Adam leads Taboola’s business strategy, execution and development while maintaining the company’s unique culture of quantification and innovation.<br><br>Prior to founding Taboola, Adam developed his analytical skills while serving as an officer in an elite mathematical unit of the Israeli National Security Agency for near 7 years (don’t ask, he can’t tell you).<br><br>Adam is an honored alum of the Israeli Defense Forces’ elite Mamram computer science training program, graduated first in his class at the Officers Academy of the IDF, and was recently named one of Israel’s top “30 under 30” business managers by TheMarker.<br><br>Adam authors a column for Mediapost on the video industry and is a regular speaker for Steaming Media, NAB, NewTeeVee, ELEVATE, Meetup, The Israel Conference, MIT (Sloan) , Bloomberg West, and others<br><br>Specialties: I build things
Explore more posts
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content