Notes from
the Build

Voice AI Open Models India October 27, 2026

Maya 2 Native: An Open Voice Model Built for the Languages India Speaks

Serverless Ventures

A five-person Bengaluru lab released an open, Apache-2.0 speech model for eleven Indic languages and English, with cross-lingual voice cloning, and it ranks second for Hindi among real-time models on VoiceArena. The founders' bet, how the model works, its benchmarks and serving numbers, where it struggles, and what to check before you deploy it.

Read the analysis · ~9 min read

AI Video How-To AI Media October 26, 2026

How to Get the Most From LTX-2.5: Prompts, Settings and Cost for AI Video

Serverless Ventures

LTX-2.5 writes the picture and the soundtrack in one pass and invents whatever you leave out. How to prompt it like a shot list, when to use fast or pro, which settings matter, why ten drafts on fast cost less than one 4K final on pro, and how to run the open weights on your own GPUs.

Read the guide · ~10 min read

AI Infrastructure Inference MLOps October 26, 2026

Ship the Kernels With the Weights: Compiled GPU Artifacts Are Build Outputs

Serverless Ventures

NVIDIA's ModelExpress loads DeepSeek-V4-Pro weights in 11 seconds, yet startup only fell from 8 minutes to 7. Shipping the compiled kernel caches took it to 1m 44s. Why compiled GPU artifacts should be versioned, environment-specific build outputs, the seven-step pipeline that follows, and how to get the cache key right.

Read the analysis · ~9 min read

Market Research IPO Analysis AI Labs October 26, 2026

Reading Anthropic's S-1: The Numbers, the Bear Case and the Questions It Leaves Open

Serverless Ventures

We read the filing. Revenue went from $41M in 2023 to $16.3B in the first half of 2026, compute fell from twice revenue to half of it, and the $80.7B net loss is mostly a non-cash charge. The risk sits in compute contracts that run to 2036 and a structure that leaves public shareholders without control. The strongest bear and bull takes, and ten questions worth sitting with.

Read the analysis · ~12 min read

Physical AI Edge AI Industrial AI October 26, 2026

Physical AI Lives Inside the Fence: Why It Needs the Edge, Not the Cloud

Serverless Ventures

Physical AI looks less like a humanoid and more like a server rack in a pulp mill. The control loop stays put; the judgment layer above it moves inside the fence, for data sovereignty, security, uptime and a token bill that grows with every asset. What the Schneider Electric, NVIDIA and HPE edge agents change, and which roles on the floor shift first.

Read the analysis · ~8 min read

Fine-Tuning LLMs AI Engineering October 25, 2026

LoRA and Beyond: Five LLM Fine-Tuning Techniques, Explained

Serverless Ventures

Full fine-tuning a 7B model needs over 100GB of GPU memory. LoRA learns the change instead, with two thin matrices. LoRA-FA freezes half of them, VeRA shares them across layers, Delta-LoRA lets the frozen weights move, and LoRA+ trains one half faster. Each one drawn out, with its paper's numbers, plus QLoRA, DoRA and how to choose.

Read the explainer · ~9 min read

AI Infrastructure Data Centers Market Outlook October 25, 2026

Stranded Compute: Why 2027 Could Be AI Infrastructure's Clearing Year

Serverless Ventures

GPU rental prices are up and compute is about to be stranded, both at once. Chamath Palihapitiya's view from the energized-land side: thinly capitalized neoclouds, slow hyperscalers, and pipeline delays that will catch people off guard in 2027. We reconcile the paradox, check the Anthropic S-1 angle, map who wins and loses, and list the questions to ask any compute provider.

Read the analysis · ~8 min read

Open Models Local AI AI Agents October 25, 2026

What a 4GB Model Is For: NVIDIA Nemotron 3.5 Bolt E4B

Serverless Ventures

NVIDIA retrained Gemma 4 E4B for instruction following and tool calling and fit it into about 4GB of GPU memory. It beats Qwen 3.5 4B at obeying system prompts, emitting valid structure and knowing when not to call a tool, and trails it on long multi-step tasks. That makes it the worker, not the brain: what the 4GB really means, the benchmarks, and a test plan.

Read the analysis · ~7 min read

Startup Trends Market Research AI Companies October 14, 2026

LinkedIn Top Startups 2026: Smaller Teams, Bigger Bets

Serverless Ventures

We read all 50 entries on LinkedIn's 2026 U.S. list and counted. 28 sell AI as the product, and 17 of the 21 companies founded since 2023 are AI-native. The median team is about 190 people, capital flows to compute rather than payroll, the agents sell completed work, and the No. 2 company sells expert human judgment to AI labs. Plus what the list measures, and what it misses.

Read the analysis · ~8 min read

Physical AI Robotics Data AI Infrastructure October 13, 2026

The Five Muscles of Physical AI: What It Takes to Scale Robot Data

Serverless Ventures

Language models inherited the internet's text. Robots inherit nothing: the data has to be captured before it can be learned from. Sequoia's note on backing Mecka names the five capabilities that takes (research, hardware, operations, software and deployment) and argues the moat is doing all of them at once. We unpack each one, map it to Leo Su's robotics data farm framework, and list eight questions to ask any robotics data company.

Read the analysis · ~8 min read

Growth Strategy Product Metrics AI Products October 3, 2026

AARRR, Revisited: Pirate Metrics for AI Products

Serverless Ventures

Dave McClure's pirate metrics reduce a startup to five numbers: Acquisition, Activation, Retention, Referral and Revenue. The value is not the acronym but the discipline of fixing the leakiest stage first, and retention before acquisition. For AI products the funnel needs three amendments: activation is the first useful output, retention is read after the novelty spike, and revenue has to clear inference cost as well as acquisition cost.

Read the guide · ~9 min read

Compound Startups AI Agents Data Infrastructure September 16, 2026

The Compound Startup Sandwich

Mohinish Shaikh

Own the data below SaaS and the agents above it, and the product surface in the middle stops being a moat. GPT-4 level reasoning fell 312x in 21 months, a product line now takes 3 to 6 months, and B2B data still decays at 2.1% a month. Salesforce renamed five lines to Agentforce and four of them grew at single digits or below, while 47% of the records its data platform ingested in FY26 never moved in at all. With four charts, eight tables, a self-test, and the counter-case at full strength.

Read the analysis · ~17 min read

AI Investing Agent Evaluation Quant Research September 14, 2026

Alpha Without Prediction

Pragadeesh VS

Language models cannot forecast prices, and the honest literature stopped pretending otherwise. Three 2026 studies dismantled the field's headline results, and a live fleet of autonomous trading agents lost $217,000 across 7.5 million model calls while converging onto the same trade. What survives is not forecasting: agents do not raise your skill per bet, they raise how many times you can use it. Eleven signal families graded durable, decaying, unproven or illusory, with the six tests each one has to pass.

Read the research note · ~21 min read

AI Media Film Economics Production Systems September 8, 2026

What If a Movie No Longer Needed to Cost Millions?

Serverless Ventures, Media

A feature film is a project management problem wearing an art form's clothes. Working the numbers line by line against Hanuman Ansh, the ₹2 crore film that grossed ₹72 crore, a $10M feature falls to roughly $1.18M once the workflow is restructured and the inference bill is honestly priced. But the cost collapse is not the disruption. The disruption is that a film stops being an artifact and becomes a build.

Read the analysis · ~13 min read

Market Research RL Environments AI Data Economy September 7, 2026

Build the Environments, Not the Workforce

Mohinish Shaikh · Pragadeesh VS

A market-entry assessment of six high-value AI data categories. The commodity labeling tier is being automated toward zero; the layer above it, reinforcement-learning environments with working verifiers, is the only category where the deliverable is software, the buyer pays per artifact rather than per hour, and a two-person engineering team can compete on craft instead of headcount. With price anchors, incumbent mapping, jurisdictional constraints, and a twelve-month entry plan.

Read the assessment · ~14 min read