Neal Ghosh

Systems of records are losing their grip. We are moving toward a world where the value isn’t in storing data, but in owning the “context graph” that automates the actual work. I sat down with Ashu Garg on this episode of the podcast to discuss everything Context Graph. Here he is explaining the it's changing the game fast. Full episode link in the comments

36

3 Comments

Runtime now runs end-to-end on Mistral AI models 🇪🇺 Runtime clients can now ship full features without a single call to a proprietary model. From intent capture to pull request, including PRD creation and code generation. The entire pipeline can run on Devstral 2. On real delivery workflows, it’s competitive with flagship proprietary models, while being faster and far more efficient on tokens. Most AI coding tools compensate for missing structure with bigger models. When context is implicit, the model has to infer architecture, reconstruct domain rules, and guess invariants. That drives token burn and architectural drift. Runtime goes the other way. Before any agent runs, it operates against a Context File System built during setup. Deterministic. Governed. Explicit. When context is engineered and agents are governed, the model becomes a replaceable component. We’ve run the same pipeline on Opus, Codex, and now end-to-end on Mistral. Clients choose based on cost, latency, data residency, or capability. No re-architecture required. This matters beyond economics. Devstral 2 is a European model. Runtime is a European platform. For teams with data residency, IP sensitivity, or regulatory constraints, that’s not a preference. It’s a requirement. We’re actively benchmarking Runtime across different codebases and delivery contexts, beyond demo-style coding benchmarks. If you want to see what end-to-end agentic delivery looks like on your own stack, let’s talk 👋

18

What happens when boomers exit the market? | A clip from my Substack post, EP028: The Architect Who Became the Grey Rabbit of Finance. Watch it here: https://lnkd.in/eiT_7Dse

1

The best deep tech companies we have seen don’t just advance their science. They progress across multiple layers at once; technology, infrastructure, market readiness, and execution. When one of those lags too far behind, momentum stalls. To make this easier to assess, we created the 𝗣𝗲𝗿𝗶𝗼𝗱𝗶𝗰 𝗧𝗮𝗯𝗹𝗲 𝗼𝗳 𝗗𝗲𝗲𝗽 𝗧𝗲𝗰𝗵 𝗕𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗕𝗹𝗼𝗰𝗸𝘀—a 64-element framework we use to decode deep tech startups. It gives founders and investors a shared language. We can point to what’s in place today, what’s still in progress, and what evidence will unlock the next round of growth. It maps the entire journey: → Core technology pillars → Tech enablers & infrastructure → Application domains → Market signals & moats → Risk factors Why does this matter? When we invest, we are looking for visionary founders with groundbreaking ideas and the ability to execute. That means lighting up clusters of these elements, not just one row. A strong TRL without a go-to-market fit isn’t enough. Neither is early customer traction without deep IP, the right talent, and regulatory readiness. If you find it useful, share it with your network and save it for your next diligence.   #Venturecapital #AI #Deeptech #Startups   Follow us at APEX Ventures and subscribe to our newsletter for exclusive content on groundbreaking Deep Tech startups: 🔗 https://t2m.io/EV2qHQuo

167

4 Comments

We have a whole philosophy and process around antiportfolio companies at Pioneer Fund. When we identify an antiportfolio company, we do a full archaeological dig. Every opinion from the investment team, every message sent internally, every piece of information we had. Who was excited and why? What were we worried about? How fast did we move? We analyze and think hard. The goal isn’t to assign blame or feel bad. It’s very easy in retrospect to say “we should have invested.” There are two goals: 1) Update our mental models so that we would be more likely to make the right decision, given the info we had at the time. 2) Identify any process improvements that would have helped us. We’ve made dozens of mental model updates this way, and we’ve made some radical upgrades to our process.

29

4 Comments

The “NVIDIA of Chronic Disease” comparison is actually a strong framework. NVIDIA built the computing infrastructure for the AI boom, while Eli Lilly may be building the therapeutic infrastructure for the chronic disease era. GLP-1 drugs are unique because they impact multiple major conditions simultaneously: obesity, diabetes, cardiovascular disease, and recently neurodegeneration. That kind of platform dynamic is rare in medicine. If that thesis plays out, the addressable market could end up being far larger than what most analysts are currently modeling. ** To read article Tap X in right hand corner and read for free, no need to enter email **

1

🚨 Waymo at 600x revenue. Perplexity at 90x. OpenAI at 70x. We’re not in a normal market. We’re in an expectation-driven market, and AI is the narrative pulling it forward. In a sharp deep dive by Palle Broe and Ben Lang, 37 of the world’s most valuable private companies were ranked by revenue multiple. A metric that shows not how much money a company makes, but how much investors believe it will make. And the numbers are wild: What’s a revenue multiple, and why should you care? A revenue multiple = valuation ÷ revenue. If a company makes $100M a year and is valued at $1B, its revenue multiple is 10x. This number tells us how much investors are betting on a company’s future, how valuable its growth, IP, and market potential are relative to its current earnings. So when Waymo is valued at $45B on just $75M in revenue (600x), it’s not because of what it is, but what it might become. Key takeaways from the data: 🔹 AI is where the speculative premium lives. Waymo, Perplexity, OpenAI, and Anthropic are commanding the highest multiples of all, and none of them are pulling in massive revenue yet. The market is pricing in future dominance of foundational AI, autonomous systems, and monetizable models. 🔹 Not all revenue is valued equally. Stripe generates $5B in annual revenue and sits at an 18x multiple. Fanatics pulls in over $8B. Yet its multiple is only 4x. These are solid, operational businesses, but they’re not riding the hype curve like AI is. Investors are placing bigger bets on scalability and long-term disruption than near-term profit. 🔹 Private vs public markets: a growing disconnect. More than 90% of the companies in the study have a higher revenue multiple than public SaaS companies, which are averaging around 7x. This points to a growing gap between private market expectations and public market discipline. Valuations in the private world may still be inflated or early bets on category leaders. What this means: - For founders: Revenue growth still matters, but category, timing, and investor narrative matter more when it comes to valuation in early and growth stages. - For investors: The multiples are a signal of conviction, not just value. Expect volatility, especially in companies where revenue has to “catch up” to the price. For the market: AI is the new crypto. It commands attention, premium pricing, and inflated multiples, but long-term winners will be those who deliver sustainable business models, not just flashy demos. The full breakdown, chart, and company list is in the comments. worth a look if you're tracking private markets, AI, or early-stage investing. #AI #StartupValuations #PrivateMarkets #Waymo #OpenAI #Perplexity #Anthropic #VentureCapital #TechTrends #RevenueMultiples #StartupFunding #DeepTech #FutureOfAI #ValuationMetrics #Founders #InvestorInsights

6

1 Comment

A new investment category is forming. Most people are 𝗺𝗶𝘀𝗿𝗲𝗮𝗱𝗶𝗻𝗴 what it means. OpenAI just moved to form a ~$10B JV with TPG, Advent, Bain Capital, and Brookfield. Anthropic is doing the same with Blackstone, Permira, and Hellman & Friedman. The headline: AI labs are embedding themselves inside PE firms to deploy 𝗲𝗻𝘁𝗲𝗿𝗽𝗿𝗶𝘀𝗲 𝘀𝗼𝗳𝘁𝘄𝗮𝗿𝗲 across portfolio companies. But that's not the 𝗿𝗲𝗮𝗹 story. The real story is what happens 𝗻𝗲𝘅𝘁. When AI stops being a tool these firms buy and starts being the engine that finds, structures, and executes the deals themselves. We're moving from: → AI as 𝘀𝗼𝗳𝘁𝘄𝗮𝗿𝗲 𝗹𝗮𝘆𝗲𝗿 → AI as 𝗰𝗮𝗽𝗶𝘁𝗮𝗹 𝗮𝗹𝗹𝗼𝗰𝗮𝘁𝗼𝗿 Private market alpha has never come from picking better companies. It comes from seeing what others can't: 𝗯𝗿𝗼𝗸𝗲𝗻 𝗰𝗮𝗽𝗶𝘁𝗮𝗹 𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲𝘀. Fragmented ownership. Trapped liquidity. Misaligned investors who haven't found each other yet. That's where transactions are born. 𝗕𝗲𝗳𝗼𝗿𝗲 anyone calls a banker. Most firms are still asking: "How do we use AI to analyze deals 𝗳𝗮𝘀𝘁𝗲𝗿?" The better question: "Can AI 𝗽𝗿𝗲𝗱𝗶𝗰𝘁 which capital structures will break and 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗲 the transaction before it exists?" That's what we've been building with GIA. A system designed to: → identify 𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗮𝗹 𝘀𝘁𝗿𝗲𝘀𝘀 in private markets → predict 𝘄𝗵𝗼 𝘄𝗶𝗹𝗹 𝘁𝗿𝗮𝗻𝘀𝗮𝗰𝘁 → surface 𝗲𝘅𝗲𝗰𝘂𝘁𝗮𝗯𝗹𝗲 𝗱𝗲𝗮𝗹𝘀 before they're on anyone's radar We call it 𝗦𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗮𝗹 𝗔𝗹𝗽𝗵𝗮. Liquidity investing powered by predictive private market intelligence. 𝗡𝗼𝘁 waiting for exits. 𝗖𝗿𝗲𝗮𝘁𝗶𝗻𝗴 them. Everyone's talking about AI in private markets. Almost no one is asking the 𝗿𝗶𝗴𝗵𝘁 question yet. What do you think the winners look like? The AI labs, the PE giants, or 𝘀𝗼𝗺𝗲𝘁𝗵𝗶𝗻𝗴 𝗲𝗻𝘁𝗶𝗿𝗲𝗹𝘆 𝗻𝗲𝘄?

9

1 Comment

SVB/Carta(Peter Walker) posted some new data on benchmarks for Series As. * Median ARR is $2.8M * Top quartile ARR is $6.9M A few observations: * Absolute value of ARR matters but quality of ARR matters even more. It is not just about reaching $2.8M in revenue but other things matter. Primarily growth rate and time it took the company to get there. Bessemer Venture Partners put out a great article defining SuperNovas (companies that get to $40M ARR in one year) and Shooting Stars (companies that get to $3M ARR in one year). Other things that impact revenue quality margins, acquisition costs, expansions, etc.). * Getting to Series A is still quite hard (2x harder than for companies started in 2020, 2021) ~10% of companies raise series A within 12 months ~20% of companies raise series A within 24 months This is still 2x lower than in 2021 and 2022. * Founders: The bar for Series A is higher. It is no longer $1M ARR. If you want to raise from top firms, you are going to get compared to the very top companies - the Super Novas and Shooting Stars. Maybe it took you a while to get to figure it out (get to $500k- $1M ARR) but if you are out trying to raise Series A, you need to show that you are in the Top Quartile growth trajectory. You don't need to be at the $3M ARR mark  but you need to show that you are on that path (at least two quarters of at least 20% growth m-o-m is 3x in 6 months) and you can do that efficiently. And, we are talking real ARR (vs cARR, pilot ARR). My friend Carlotta "Lotti" Siniscalco has a great post on this. * Note that there are always exceptions on either side: 1) companies with <$1M ARR, raising a competitive Series A ($2.8M is the median not the bar) - sometimes team trumps all metric rationale and 2) companies with >$2.8M ARR not raising Series A - can be a limited market, a team issue , etc.

89

5 Comments

From Palantir's earnings call on AI: “As AI continues to relentlessly advance, the market has become increasingly aware of the most important and fundamental technical reality: LLMs, on their own, are at best a jagged intelligence divorced from even basic understanding. In one moment, they may appear to outperform humans in some problem-solving task. But in the next, they make catastrophic errors no human would ever make."

19

Why was the Fivetran-dbt merger all but inevitable? Fivetran & dbt Labs announced their merger yesterday. The all-stock deal combines two companies into an entity approaching $600 million in ARR. The beauty of the modern data stack was the explosion in choice. As the cloud exploded onto the scene, the legacy data warehouse was replaced by a collection of fast-moving platforms. In that era, specialization won. The pendulum is now swinging back towards consolidation. Why? The answer lies in compute economics & revenue scale asymmetry. The table below shows why. There are three different categories of software within this subset of the ecosystem: 1. Ingestion takes data from software & moves it into a cloud data warehouse. Snowflake acquired Datavolo, which commercializes the open source product Apache NiFi, calling it Openflow. Databricks acquired Arcion for ingestion through change data capture, calling it LakeFlow Connect. Fivetran focuses exclusively on this layer. 2. Transformation means reformatting the data within the cloud data warehouse. Snowflake launched native dbt Projects on Snowflake. Databricks offers Delta Live Tables, native SQL, & Python, plus supports hosted dbt through Databricks Workflows. dbt Core/Cloud is the leading independent transformation tool. 3. Compute revenue is generated when we ask questions of our data. Snowflake remains one of the leaders in structured data analysis with their cloud data warehouse. Databricks’ compute is their own as well. Here’s the asymmetry in one number. Compute represents 72% of the overall market ($7.6B of $10.6B). As a result of their massive operations, Snowflake & Databricks exert significant gravity within the ecosystem. They have expanded beyond the compute market to impose their presence & capture marginal revenue within customers, pressuring the competitive ecosystem. That’s not to say these components are independent. George Fraser analyzed Snowflake workloads in September 2024, finding transformation represents 40-45% of total Snowflake compute, which means even at smaller scales, startups can have significant impact on these behemoth businesses. The Fivetran-dbt merger is an inevitable evolution of a maturing market. Two unicorns must partner to compete against two decacorns. They solve two of the three customer problems. But not yet compute. One could surmise this consolidation signals the end of the modern data stack. I view it differently. The MDS has succeeded beyond our expectations. The stakes are higher now. Broad platforms, fast growth, & AI-native architectures define the next phase. Expect more consolidation. Category revenue estimates based on public disclosures & company filings. Ingestion: Informatica ($1.64B FY2024), Fivetran ($300M est.), Talend ($350M), others ($200M est.). Transformation: dbt Labs ($300M est.), others ($200M est.). Compute: Snowflake ($3.6B FY2025), Databricks ($4.0B ARR Aug 2025).

42

9 Comments

The unit economics of autonomy will be won or lost in the Infrastructure Orchestration Layer. During the Evercore Panel, we discussed why scaling a fleet is a distributed systems problem disguised as a logistics one. 1. The Multi-Depot Mesh 🛰️ In any major market, you aren't running one "garage." You're managing multiple facilities with heterogeneous capabilities. The challenge isn't just routing a car to "a" depot; it’s a global optimization problem across the entire mesh. Where do you route for sensor calibration vs. deep cleaning vs. high-throughput charging to maximize fleet-wide uptime? 2. CDNs for Physical Autonomy 🌐 We need to stop thinking about depots as parking lots and start thinking of them as Physical Content Delivery Networks (CDNs). The Internet CDN: Delivers bits to the edge to reduce latency. The Physical CDN: Delivers high-uptime vehicles to the edge to meet localized demand. How do we build the logic that determines how vehicles (our "packets") move through this network to provide the best performance at a market level. 3. Load Balancing the "Electron Demand" 🔋 Physics is the ultimate constraint. As charging windows shrink (higher kW, larger packs), the scheduling of secondary workflows: cleaning, preventative maintenance, etc. becomes a high-stakes bin-packing problem. We're solving for the supply and demand of electrons over a rolling 24-hour cycle across an entire urban grid. #Engineering #DistributedSystems #AV #Infrastructure #TechDeepDive #Evercore

123

3 Comments

When Sara Beykpour started building Particle in early 2023, there was no playbook — no dashboards, no eval tools, no prompt libraries. Two years later, the AI stack looks entirely different. In our latest piece from our Adverb Ventures new series “Field Notes,” Sara Beykpour and Jennifer Jang share how they’re building their LLM stacks, what to buy, and what actually gives them an edge. They get into: 🧰 Build vs. buy → control vs. speed ⚙️ Prompt & eval design → how to engineer for reasoning 🧩 Evals vs. A/B tests → keeping your baseline solid while you climb higher It’s a candid look at the trade-offs founders are making as they scale AI products in real time. 👉 Read it here: https://lnkd.in/gCUXgEQv

51

"Do we really need RL when SFT can match performance with just high-quality curated data?" - a question that keeps popping up as new techniques emerge. A recent paper settles this debate with a crucial insight: it's not about performance on the target task, it's about what happens everywhere else. They trained SFT and RL models on identical math datasets, but only RL models improved at non-math tasks like conversation and instruction-following. The culprit? SFT causes substantial representation drift in the model's internal geometry, while RL preserves the underlying feature structure. Using PCA analysis and token rank shift measurements, they show SFT induces large distributional changes that hurt generalization. Their Transferability Index consistently shows RL models transfer better across domains, while SFT models often show negative transfer - actually getting worse at non-reasoning tasks. The takeaway: If you're fine-tuning for narrow, domain-specific tasks, SFT might be sufficient. But if you need models that maintain their broader capabilities while learning new skills, the evidence strongly points toward allocating more resources to RL approaches. Sometimes the question isn't "can we do it cheaper?" but "what are we breaking in the process?"

12

1 Comment

2) 2026 setups coming into focus: Heading into 4Q EPS, average CY26 total revenue estimates are effectively unchanged vs. pre-3Q25 (-0.2% on average), though TTAN/VEEV/GWRE/WK all saw a >1% increase in estimates. Over the same period, CY27 total revenue estimates saw modest downward revisions (-0.8% on average), with only TTAN/VEEV/GWRE seeing positive revisions in our coverage (Exhibit 8). When assessing 2026 setups across our coverage, we note that Consensus estimates currently imply a ~2.5pt deceleration in YoY growth on average vs. 2025 Consensus growth. Across our coverage, only BL/WAY Consensus estimates imply an acceleration in growth off of 2025 levels (though we note 2026 estimates for WAY include inorganic contributions from the company's acquisition of Iodine Software, Exhibit 9). On 2026 total revenue estimates, we are most above Consensus for TTAN/WAY/VERX/GWRE and most below Consensus for VEEV/SPT/NCNO. For CY27, we are most above Consensus for TTAN/WK/NCNO (Exhibit 10). While we anticipate management teams across our SMID-cap software coverage universe will maintain a prudent approach to FY26 guidance (a continuation of the trend we observed throughout FY25), we expect investors will increasingly focus on the magnitude of quarterly beat-and-raises to get more positive on shares against what we view as effectively de-risked 2026 estimates.

1