Google's Cloud Revenue Converges to NVIDIA's Growth Rate

Google Cloud posted 82% year-over-year revenue growth to $24.8 billion in Q2 2026, significantly exceeding consensus estimates. Notably, its growth curve now mirrors NVIDIA's Data Center segment — both driven by surging demand to train and run AI models, one through compute rentals, the other through chip sales. Google Cloud's backlog reached $514 billion, up 385% year-over-year, with over half expected to convert within 24 months. Operating margins expanded to 35.6%. Across the four major cloud hyperscalers, tota…

7/22/2026

AI Engineering Productivity is Anything But Normal

A Redpoint Ventures analysis segments AI-driven engineering productivity gains into three distinct tiers. Companies that simply distribute AI coding tools see roughly 20–30% improvement, sometimes accompanied by rising bug rates. Those building agentic orchestration layers integrated with tools like GitHub and Slack report over 3x gains, with NVIDIA and Amplitude cited as examples. A third tier leveraging software factory approaches via tools like Cognition's Devin or Factory.ai reports over 8x efficiency improvem…

7/21/2026

Open Models Tack Toward the Frontier

A Redpoint Ventures analysis finds open-source AI models rapidly narrowing the performance gap with closed models. Since DeepSeek R1 in 2025, the two camps have run nearly in parallel, with GPT-5.2 and Fable 5 pulling slightly ahead in 2026. A fresh wave of open-weight releases followed, including Moonshot's Kimi K3 (2.8T parameters), Alibaba's Qwen 3.8 (2.4T), and DeepSeek V4. Inference costs for the median open-weight frontier model run roughly 15% below GPT-5.2, while the cheapest option, DeepSeek V4 Flash, is …

7/20/2026

The Blank Slate AI Strategy

Redpoint Ventures draws a parallel between Slate Auto's stripped-down electric pickup truck and Thinking Machines Lab's open-source 975B-parameter model Inkling, arguing both embody a blank-slate strategy built for customization. Inkling is trained on 45 trillion tokens and released free via GitHub, but monetized through Tinker, the company's paid fine-tuning platform. Redpoint sees this model, ship a general-purpose base and charge for customization, as an emerging commercial framework for open-source AI, analogo…

7/16/2026

The Harness Is the New Battleground

Satya Nadella and Alex Karp both warned in the same week that enterprises effectively pay for AI twice: once in fees, and again by exposing proprietary data to frontier labs. The concern moved from theoretical to concrete when a researcher found xAI's Grok Build silently uploading developer codebases to xAI's cloud. Redpoint Ventures argues that enterprise AI harness-layer tools like Claude Cowork and Cursor will compete primarily on how well they balance productivity with data protection. Meanwhile, startups gene…

7/14/2026

The AI Colander

Analysis from Redpoint Ventures highlights rapid commoditization in the frontier AI model market. The average leading model retains its top position for only 41 days, with user retention rates at five months ranging between high single digits and roughly 40% — closer to mobile games than traditional SaaS. Benchmark evaluation is also shifting: cost efficiency, measured in tokens and dollars per task, now rivals raw performance. Microsoft's MAI-Code-1-Flash matches Claude Haiku 4.5 on SWE-Bench Verified using 60% f…

7/13/2026

Three Years In

Theory Ventures marks its third anniversary by reflecting on how AI has reshaped software and venture markets faster than anticipated. The firm's founding thesis — that AI would transform how software is built, sold, and operated — has materialized at an accelerating pace, with frontier models releasing every 41 days on average. Key structural shifts include the collapse of traditional funding stage definitions, seed rounds now ranging from $1 million to $500 million, inference emerging as the dominant AI market, …

7/10/2026

8.9 Million AI Users

7/9/2026

The AI Preflight Check

Redpoint Ventures' Tom Tunguz argues that AI agent performance hinges on memory architecture, not raw context size. His 'preflight' approach mirrors a pilot's pre-takeoff planning: before executing a query, the agent scans a skills library and loads only relevant skills into the context window. A local Ornith 35B model handles roughly 80% of routine tasks, while harder tasks route to frontier models. A watchdog logs every skill invocation, and overnight asynchronous inference refines the skill set—turning repeated…

7/8/2026

The $10B FDE Boom

Over the past 12 months, AI companies have committed $9.75 billion to forward-deployed engineering (FDE) — embedding engineers inside customer organizations to implement AI. Once a Palantir signature, FDE has become an industry default. Three structural models are emerging: the Balance Sheet model (Microsoft, Amazon fund teams internally), the Standalone model (OpenAI raised $4B at a $14B valuation; Anthropic raised $1.5B from Blackstone, Hellman & Friedman, and Goldman Sachs), and the Partner Ecosystem model (Goo…

7/7/2026

AI Worldviews

The Economist administered the World Values Survey to 25 frontier AI models, mapping their positions on axes ranging from traditional to secular and from survival-focused to individualist self-expression. Most models clustered toward self-expression, consistent with their English-dominant training data. Notably, GPT-4o and DeepSeek R1 showed near-identical value profiles despite originating in San Francisco and Hangzhou respectively, while DeepSeek R1 and DeepSeek V4 Flash diverged sharply on the secular-tradition…

7/6/2026

Search

Search the tomtunguz.com archive of essays covering topics such as AI, SaaS, startups, and venture capital to find relevant insights and analysis.

7/3/2026

Most AI Work Can Wait

A router-first approach to AI agent architecture is gaining attention as a cost-reduction strategy. Rather than selecting a model upfront, teams should design a routing layer that classifies tasks and directs them to the appropriate model tier. This approach can route 70-80% of traffic to free local models or low-cost async models, reducing AI spend by over 90%. Coinbase CEO Brian Armstrong cited similar routing and caching defaults as key to keeping AI costs flat even as token usage grew significantly.

7/1/2026

The CIO's Choices are Clear in 2026

Analysis from Redpoint Ventures highlights a sharp divergence in public software market performance heading into 2026. Infrastructure and Dev Tools lead all sectors with 68.5% one-year gains, driven by names like DigitalOcean, Datadog, and Cloudflare. Security follows as the only other positive sector at plus 17.6%. Meanwhile, productivity and collaboration software including Monday.com, HubSpot, and Atlassian has struggled. The pattern signals that enterprise CIOs are concentrating budgets on AI infrastructure an…

6/30/2026

When AI Costs More Than the Engineer

A Redpoint Ventures analysis highlights a striking shift in software cost structures: Anthropic spends roughly 2.3 times its total payroll on compute, amounting to approximately $2 million per employee annually. The top one percent of non-frontier companies spend around $89,000 per engineer on AI, while the median company spends just $137. Three scenarios—bear, base, and bull—project how quickly the broader market might close this gap by 2029, with bull-case estimates suggesting AI costs per engineer could exceed …

6/29/2026

What If There Is No Moat Yet?

Redpoint Ventures argues that founders should distinguish between two types of competitive moats. Leading moats exist at founding—proprietary datasets, novel architectures, technical differentiation—and are essential for infrastructure companies that require heavy research and capital. Lagging moats, by contrast, are earned over years through scale, brand, channel relationships, and embedded workflows. Application-layer companies like Salesforce typically win with lagging moats. The key insight: founders should co…

6/26/2026

Full Sail on Asynchronous Inference

Theory Ventures argues that current AI inference infrastructure is optimized for real-time, low-latency interactions, creating unsustainable costs when running multiple agents in parallel over extended periods. Through internal experimentation with queued, asynchronous workloads, the firm encountered Sail Research, co-founded by Neil Movva and Samir Menon. Sail routes requests across open models such as DeepSeek, Qwen, Kimi, and GLM to minimize per-token costs, positioning itself in the batch-tier segment of a seg…

6/25/2026

Defending Against AI-Powered Attackers

Redpoint Ventures' Office Hours series will host Sunil Agrawal, CISO at enterprise AI search company Glean, on July 9 for a focused discussion on security readiness in the age of AI-driven attacks. Topics include how AI compresses attacker reconnaissance timelines, the erosion of traditional phishing detection signals, deepfake threats to approval and payment workflows, and the organizational changes security teams must make. The session runs 15 minutes online in a live call-in format with no prepared slides, desi…

6/24/2026

The Quietest Part of Startupland isn't so Quiet

Despite venture funding in crypto sitting at multi-year lows, structural developments suggest the sector is far from dormant. Stablecoin issuers now hold $165 billion in U.S. Treasury bills — roughly 2.5% of the total T-bill market — ranking them behind only Japan among foreign holders. Meanwhile, Hyperliquid, a four-year-old decentralized exchange, has reached a native token market cap of $59 billion and is generating annualized trading fee revenue of approximately $610 million. Real-world assets including gold, …

6/23/2026

So You Want to Sell Inference

6/22/2026

Databricks Widens the Lead on the Yellow Brick Token Path

6/17/2026

5x for Free : The Local Coding Stack

A Hacker News thread reveals the maturing local coding stack: Qwen 3.6 35B-A3B dominates at 33%, with Pi leading as agent harness at 49%. Local models offer 5x speedup, privacy, and zero cost, approaching Claude Sonnet 4.6 benchmark scores.

6/16/2026

The Golden Age of AI Applications

6/15/2026

A CEO's Cost of Capital Advantage

6/12/2026

The AI Glass Ceiling

Anthropic's Fable model sets a new AI performance ceiling with strong guardrails. Despite restrictions, it delivers remarkable results—Stripe migrated 50M lines of Ruby in a day. AI techniques evolve rapidly, and controlled rollout is essential for stability.

6/10/2026

The Substitution Wave in AI

Three forces are reshaping AI cost structures: labs moving into apps, rising frontier model prices, and capable open-source models. Companies like Coinbase, Lindy, and Harvey are substituting expensive closed models with cheaper alternatives, reinvesting savings into greater token usage and intelligence.

6/7/2026

The Minimill of AI

A laptop handles 78% of AI tasks locally, routing only complex ones to the cloud. This two-lane system boosted throughput 25%, cut task duration from 47s to 19s, and slashed queue age by 94%, mirroring Nucor's disruptive minimill model.

6/5/2026

Local in Reality

This article explores what it truly means to be 'local' in reality, examining the concept beyond surface-level definitions and its implications in practical contexts.

6/4/2026

Intelligence Per Dollar

Microsoft introduced 'average token usage' as a new benchmark metric, signaling a shift from raw AI performance to cost-efficiency. Companies like Uber and Salesforce face budget overruns, pushing the industry toward pricing by outcome rather than token consumption.

6/3/2026

The Thriving Ecosystem of Open Models

Open-weight models now generate 69.1% of token volume on OpenRouter, overtaking closed models. Rapid competition among DeepSeek, Qwen, Arcee, and others drives innovation, with developers increasingly routing production traffic to open models.

6/2/2026