We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
1. Funding & Deals
Fireworks attracted a $10M conviction check around self-owned specialized intelligence. A current investor post says, “We wrote a $10M check into Fireworks in 10 mins,” citing Lin and Dmytro’s expertise and the thesis that future companies will build specialized intelligence on their own models and data, with Fireworks helping them do so. The useful read is founder quality plus infrastructure thesis—not a valuation benchmark: the post is a conviction signal rather than a disclosed stage-and-terms financing comp.
2. Emerging Teams
Display.dev has an early traction signal, but its ICP is still unsettled. The founder says the product reached 1,500 users three months after its May launch and continued gaining weekly usage and new users even after similar products—including Claude Artifacts—appeared weeks later. Specific search ads, personal LinkedIn/X posts, and AI recommendations drove signups; broad-match ads produced low-quality signups and newsletter sponsorships generated traffic but few users. The product is a collaboration layer for teams iterating on agent-generated documents, but the founder says its horizontal market makes the best company or user type difficult to identify. For an investor, the combination is promising distribution and retention evidence with a still-open positioning question.
A second signal is the falling cost of becoming a maker. A builder reports $24K in revenue from a weekend Replit project built on an iPhone, while a related update reports 53,000 pixels sold through the app. That is evidence of rapid platform-enabled monetization, not yet evidence of durable retention or a scalable company; the diligence question is whether this velocity generalizes beyond novelty-driven launches.
3. AI & Tech Breakthroughs
The agent stack is being repriced below the model. Parallel launched Fast at $1 per 1,000 queries, describing it as 5–10x cheaper than other APIs and 10x cheaper than the default search bundled with frontier models; it says Fast nearly matches its Advanced tier and cites an Artificial Analysis comparison of 12 search APIs across quality, cost, and speed. Parallel further claims search accounts for less than 12% of total agent cost with Fast, versus 48% for Brave and Exa and 68% for Tavily, producing claimed end-to-end savings of 2.2–2.79x. If replicated, search becomes a meaningful cost-control and routing layer for high-volume agents as model prices fall.
Small specialist models are challenging scale as a proxy for bounded tasks. A community test of webAI’s 3B TwIL-LM3 says it beat gpt-oss-120b on four of five formal-reasoning tasks, while losing the broader loose-match aggregate, 0.4488 to 0.5192. Its throughput tests were 32.9 versus 12.6 answers per second, with 40x fewer parameters and support for 4GB VRAM or CPU. The tester explicitly limits the result to narrow formal reasoning rather than general capability; the investment signal is that owned schemas and constrained workflows may justify specialist-model economics even when general benchmarks still favor larger systems.
Security and factual reliability are moving into the product layer. Claude Security scans now run on Anthropic’s Mythos 5 in public beta for all Claude Enterprise customers, with no separate model access required. Martin Casado’s framing—third-party/API, first-party Claude Code, then “no party”—captures the competitive direction: security analysis is being embedded in the coding environment rather than left entirely to an external tool.
A separate community analysis of OpenAI evaluations reports higher hallucination rates for o3 than o1 on PersonQA, and 51% for o3 versus 79% for o4-mini on SimpleQA; it cautions that these results do not generalize to every model or task, but argues that additional reasoning can elaborate a bad premise when context is weak. Its proposed production response—context-sufficiency gates, provenance, evidence-linked answers, abstention, and human review—is an architecture and governance requirement, not merely a model-selection problem. The adjacent security risk is contextual authorization: a current roundup reports that a legitimate n8n workflow became a route to remote code execution even though the trusted steps ran as designed, exposing the gap between “may this principal call this tool?” and “should this sequence run in this context?”
4. Market Signals
ServiceTitan’s cutoff of Podium is a concrete example of AI compressing platform switching costs. SaaStr reports that ServiceTitan gave roughly 1,000 shared customers about 30 days’ notice that Podium’s integration would be shut off after nine years, while ServiceTitan said Podium declined certification under its new post-AI terms. The conflict is strategic, not merely contractual: Podium’s agent business reportedly went from zero to $100M ARR in under 24 months, and its new FSM replaces scheduling and dispatch software with a 14-day migration of contacts, job history, and price books.
ServiceTitan’s API terms now bar AI systems from independently choosing endpoints, data modifications, or actions; calls must remain inside predefined certified operations, AI use must be disclosed, and the platform may require human authorization for writes. The practical distinction is between bounded workflow automation and an autonomous operator. This is not simply an incumbent losing relevance: the same report gives ServiceTitan quarterly revenue of $268.8M, up 25%, net retention above 110%, and says locations using its Max AI product more than doubled while free cash flow remained negative because of spending on Max and inference.
Investment read: treat systems-of-record integrations as leases, not assets. Model notice periods, shared-customer migration, and the probability that an open API becomes more restrictive as a partner’s agent moves toward the platform’s core workflow.
“We can build custom software fast” is becoming a weak standalone pitch. One founder reports spending about $2,000 on ads for days-or-weeks custom builds and receiving essentially no response, despite believing the speed improvement is real. The surrounding discussion says everyone can go faster now, while business buyers need proof, domain credibility, or quantified outcomes—especially in regulated work—rather than another layer over Claude Code. For early-stage teams, speed is increasingly table stakes; the moat has to be workflow ownership, trust, or evidence of a specific business result.
5. Worth Your Time
- Read — Introducing Parallel Search Fast. The clearest current articulation of search as an agent-cost and routing layer.
- Read — When a 9 year partnership breaks down in the Agent Era. A useful case study in platform dependence, agent-specific API restrictions, and switching-cost compression.
- Read — Are we paying a “Reasoning Tax” for smarter AI?. A practical, if community-sourced, framework for context gates, provenance, abstention, and evidence-linked outputs.
- Read — TwIL-LM3, 3B formal reasoning specialist. A bounded-task benchmark result that makes the specialist-versus-generalist tradeoff concrete.
Martin Casado, the a16z GP behind the firm's investments in Cursor and OpenRouter, gave an interview days after SpaceX closed the $60B Cursor deal and Stripe agreed to buy OpenRouter .
His core thesis is that AI has made capital more productive than ever: a famous, widely used model was built by about 20 people at a cost of roughly $2B+ — 'in the history of engineering efforts, we've never been able to have 20 people productively use $2 billion' — and the major story of this wave is applying large amounts of money productively in short amounts of time . He calls AI the first technology where you can 'put in $10 and reliably get something back,' versus the old pattern of engineer, wait two years, cross your fingers .
The interview agenda maps his current market views: the case both for and against 'the labs winning everything' ; models proved much stickier than assumed and real model routing is an AI-complete problem ; 'RSI' is the wrong term — the dynamic is autocatalytic ; supply constraints ease in 2028 under his 80/60 split ; what happens when subsidies stop ; AI broke marketing — users can now be bought for a dollar, Chinese operations arbitrage $200 subscription plans, and the CMO is turning into a CFO ; more private capital grows the TAM rather than inflating it, while balance-sheet investors keep misreading these companies . He also discusses why Cursor iterated faster than anything he's seen outside an Elon company, and what both deals say about strategic value versus business quality .
Parallel (@p0) — Parag Agrawal's AI web-search startup — launched Fast mode for Parallel Search, built on the thesis that agents will use the web far more than humans and need far cheaper search infrastructure. The API is priced at $1 per 1,000 queries, claims to be 10x cheaper than the default search bundled with frontier models and 5-10x cheaper than other search APIs, and is optimized for cost-efficient models (GPT-5.6 Luna, DeepSeek V4 Pro, MiniMax M3, Qwen3.8 27B).
Parallel says Fast nearly matches the SOTA accuracy of its Advanced tier and, per an independent Artificial Analysis Search Index measuring 12 search APIs, leads across quality, cost, and speed — defining the quality-vs-cost and quality-vs-latency Pareto frontiers.
For many use cases web search is a large majority of end-to-end agent cost; Parallel claims search is <12% of total cost with its API versus 48% for Brave Search, 48% for Exa Search Fast, and 68% for Tavily Search Basic, making end-to-end agent work 2.2x, 2.35x, and 2.79x cheaper respectively.
Market context: cost-per-intelligence fell 8.5x for models scoring ≥60 Intelligence Index (Fable 5 to GLM-5.3 Max) and 12.5x for ≥50 (GPT 5.4 xhigh to GPT 5.6 Luna Max); OpenAI cut GPT 5.6 Luna pricing by 80%, and DeepSeek V4 Flash, Qwen3.8 27B, and MiniMax M3 lead usage on OpenRouter and OpenCode.
Garry Tan posted that "the datacenter discourse is a documented psyop and they're winning," linking to a thread by @garryslist . The thread cites two reports (by @SamLyman33 and @bitcoinpolicy) documenting a coordinated foreign influence campaign against American AI — via CCP state media, a Shanghai-based Marxist's nonprofit network, and foreign billionaire dark money that has funneled $2B+ into US advocacy infrastructure — concluding that "AI doomerism isn't as organic as it looks" .
Martin Casado, a16z GP behind investments in Cursor and OpenRouter, argues AI is the first technology where you can put in $10 and reliably get something back, unlike earlier tech that required engineers, two-year waits, and luck. The interview was recorded just after a $60B Cursor deal and Stripe's agreement to buy OpenRouter.
Casado credits Cursor's fast iteration to founders spending 30–40% of their time on hiring and culture, and to treating software development as a product problem rather than a research problem: "They always kept the main thing the main thing, which is changing how you write software... The model's very, very important, but it was a product problem." He contrasts this with research-heavy AI companies, adding: "There's actually not a lot of companies that are actually focused on product... There's a lot of research companies... But they were like, 'We are a product company.' And I think that's actually a key differentiator in this era."
Starcloud raised $250M at a $2.3B valuation to build AI compute capacity in orbit; Manhattan West led the round, with new investors NVIDIA, Cisco, Cedar Capital, Goanna Capital, and Standard Cap, plus continued backing from Benchmark, EQTVentures, Soma Capital, NFX, and SevenSevenSix . Starcloud is scaling production: next week it moves into a 100,000 sq ft facility to reach a production rate of 100 satellites/week . Dalton Caldwell (Standard Cap) welcomes Starcloud into Standard Cap's community, calling it one of the most ambitious startups, signaling Standard Cap's conviction in the space-AI-compute thesis .
Paul Graham tested ChatGPT's map generation by asking it to map Saracen-controlled territories on the Italian mainland and reports it is "not very good at making maps yet," noting the odd contrast with broader AI worries .
Garry Tan (YC President) posted 'Legalize building societal infrastructure again' , amplifying sdamico's argument that data centers are 'the first thing that, via market forces, broke the consensus to freeze the built world in amber (decided in the mid-late 1970s)' and that the 'singularity' may be 'allowing people to resume building stuff like they could pre-1978' . This signals an emerging investment theme around regulatory tailwinds for infrastructure and data-center construction.
Y Combinator spotlighted @steipete's builder stack progression: cli → app → web (multiplayer agent sessions), signaling an emerging product paradigm for agentic AI . The linked conversation (Raphael Schaad with Peter Steinberger) covers agentic engineering concepts ("Loops, Graphs, and Florps"), separating personal vs. work agents ("Outie vs. Innie"), and startup ideas for 2026 .
Martin Casado (a16z GP behind investments in Cursor and OpenRouter) argues AI is the first technology where capital directly converts into capability: "If I'm training a model and I want it to be good at X, I can create an RL environment of X, or I can go pay someone to answer questions for X, and I can turn capital into being good at that X." He contrasts this with pre-AI software where $10 in meant years of engineering with uncertain payoff — capital went from "lottery ticket" to "vending machine," though he notes you don't know what gets worse on the other side.
In an interview recorded after SpaceX closed the $60B Cursor deal and Stripe agreed to buy OpenRouter , Casado presents both cases for and against the labs winning everything , calls real model routing an AI-complete problem , says models proved stickier than assumed , expects supply constraints to ease in 2028 with an 80/60 split , and praises Cursor's iteration speed as faster than anything he's seen outside Elon companies .
Garry Tan amplified a post arguing the "Ban data centers!" movement is a "luxury belief" policy that would have disastrous downstream consequences, adding "the luxury belief industrial complex is alive and well" . The original post contends that, love them or hate them, "the entire global economy runs on" data centers .
New research shows data centers lowered electricity rates for Dominion Energy customers in Virginia, per a post by peterjhasson linking to https://www.ethree.com/backcast-data-center-rates/. In a reply, David Ulevitch corroborates the same pattern in Nevada . This offers a counter-signal to the common concern that data center power demand raises costs for ratepayers.
- Stripe is acquiring AI model router OpenRouter; NYT reports a $7.5B price, and OpenRouter was founded by OpenSea founder Alex Atallah .
- Model routing is becoming a core AI cost/performance layer as companies optimize return on tokens ; Merge launched Merge for Workforce claiming 75x fewer tokens for better results, and Ramp launched its own Router product after buying router.com — both are Not Boring Capital portfolio companies .
- Robotics company Generalist says its GEN-1.5 is a one-shot learner, picking up tasks from a few seconds of data, simulation, watching humans, or chained smaller tasks .
- Drone delivery is scaling: Zipline and Uber partnered, with Uber investing in Zipline to target 1M deliveries/day ; Amazon is expanding Prime Air to 500 U.S. cities by year-end ; DoorDash launched DoorDash Air . A Texas customer captured an Amazon Prime Air package dropped in her pool, a reliability caveat .
A16z publicly argues "It's time to build mega data centers," claiming counties where they are already operational show more housing, higher home values, less unemployment, and more job growth since 2024 . David Ulevitch, a16z GP and former OpenDNS founder/CEO, concurs, stating "The problem with datacenters is that we don't have enough, by a very wide margin" .
Anthropic launched Claude Security scans in public beta for all Claude Enterprise customers, now running on its Claude Mythos 5 model, with no separate model access needed . Martin Casado (a16z) frames this as the rise of a 'no party' security-scanning option embedded in the first-party experience, contrasted with third-party, API-based scanners and first-party Claude Code — signaling built-in security scanning as a growing competitive front in AI coding tools .
AI/tech market signals: SpaceX is buying Cursor for $60BN, Stripe made an $8BN OpenRouter bet, Anthropic's first profit and the math behind reaching $600BN in revenue, and Lovable and Higgsfield raised mega rounds.
Founder M&A sentiment: Founders often prefer being acquired by an iconic, highly effective operator like Elon Musk over Meta, with emotional alignment and shared vision frequently outweighing brand prestige in exit choices.
Defensibility thesis for buyers: Private equity buyers should target closed systems of record with near-zero churn and predictable cash flows; closed architectures protect ecosystem budgets and create a moat against AI-agent disruption.
OpenRouter/Stripe debate: Jason Lemkin argues OpenRouter is strong for developer tools and simple chatbots, but frontier-model enterprise workflows do not rotate through 11 models, so Stripe risks owning a successful niche product rather than a platform. The counterpoint is that a premium for an elegant, deployable product can beat internal build, giving Stripe immediate AI inference flows and a second growth engine.
Hyper-growth AI investing: Overweighting early unit economics can cause investors to miss massive market waves; products scaling fast despite margin headwinds can still become category leaders and attract strategic acquirers.
Agentic-era disruption: Rapid AI improvements have broken the quarterly software roadmap; teams adopting agentic workflows can pull years of planned development forward, leaving legacy cycle-based teams behind.
M&A revenue dynamics: Strategic acquirers can price on future expansion and largely ignore existing revenue, whereas PE buyouts require precise revenue accounting.
Paul Graham argues that preventing companies from building data centers in America would not slow global AI progress — it would merely slow AI progress in America, a cautionary signal about US regulatory risk and its competitive position in AI .
Scott Kupor, Director of the U.S. Office of Personnel Management (OPM), is working with the USTechForce team to bring AI to the USAJobs federal recruitment platform, expressing enthusiasm about the initiative . This signals early-stage AI adoption within federal hiring infrastructure, a potential market signal for AI/govtech startups targeting government HR workflows.
SpaceX is buying AI coding assistant Cursor for $60BN, Stripe is betting $8BN on OpenRouter, Anthropic posted its first profit (with the math behind reaching $600BN in revenue), and app builders Lovable and Higgsfield raised mega rounds .
On the Stripe–OpenRouter deal: the bear case is that multi-model routers serve narrow developer use cases — high-reasoning B2B workflows standardize on specific models to prevent drift — so Stripe may buy a niche tool rather than a broad enterprise transaction platform . The bull case: Stripe is proven at M&A (used it to accelerate into crypto), ~5% of market cap plus cash makes buying cheaper than building, and even a niche router grants immediate access to massive AI inference flows as a second growth engine . Jason Calacanis calls OpenRouter “one of these pieces of software which is just instantly easier to deploy. It's just elegant.”
Investment signals: in hyper-growth markets like AI coding, fixating on early unit economics makes investors miss massive waves — products scaling despite margin headwinds can become category leaders and attract strategic buyers who absorb those costs . Rapidly improving AI has broken quarterly software roadmaps: teams embracing agentic workflows can pull years of planned development forward, leaving incremental-release legacy teams behind . Buyout-oriented investors are advised to target closed systems of record with near-zero churn and predictable cash flows, since closed architectures moat against disruption by third-party AI agents . Founders often prefer selling to an iconic operator like Elon Musk over entering a Meta-style corporate structure, on emotional alignment and shared vision .
Martin Casado says AI has made capital more productive than ever: one of the most popular models was built by ~20 people at an estimated cost of $2B+ . He calls it unprecedented in engineering history for 20 people to productively use $2B , and argues the major story of this wave is deploying large amounts of money productively and quickly, not just technical sophistication .
Martin Casado (a16z) says he has never worked on a VC deal that didn't involve multiple people playing critical roles in both closing the deal and company success, and argues venture is underrecognized as a team game .
