Greptile Raises $50M Series B as Perplexity, Replit and Vercel Push Local and Private AI Tooling

Key Takeaways

1. Greptile raised a $50M Series B led by Menlo Ventures at a $750M valuation, with a team of under 50 reviewing over 1% of GitHub PRs daily.
2. Perplexity Computer now runs fully locally on Mac, while Replit will support deployment into customers' own cloud environments.
3. Vercel open-sourced a fast HTML artifact builder for agents, reporting 1,000 tokens per second.
4. Exa 3.0 launched, claiming first place across all tested benchmarks including coding and GTM intelligence.
5. IREN's GPU contracts show prepayments covering 45–55% of capex, with 2026 capacity essentially sold out.
6. Tiny disclosed its $250K Poolside stake now carries an implied book value of about $11.1M.
7. Wabi 2.0 and Willow's Wally highlight a wave of collaborative and writing-focused agent products.

1. AI Infrastructure and Developer Tooling

1. Greptile completed a $50M Series B led by Menlo Ventures at a $750M valuation, with a team of fewer than 50 people reviewing and testing more than 1% of GitHub PRs daily; customers include Nvidia, Coinbase, Decagon and Vanta across thousands of engineering teams. Garry Tan argued that engineer code-change speed has risen 10x this year, shifting the bottleneck to review and testing, and that Greptile mixes frontier models with GLM, Kimi and DeepSeek to cut costs that can exceed $20 per review with frontier models alone — via 1 2 3
2. Perplexity Computer can now run entirely locally on Mac, with models and agents executing on-device without network access or sending sensitive data to the cloud, available in the Perplexity Mac app; separately, Replit will support deployment into users' own cloud environments so data does not leave local premises, aimed at government, finance and healthcare — via 1 2
3. Vercel open-sourced a fast HTML artifact builder based on libfx that gives agents persistence, with Guillermo Rauch calling the prior v0 infrastructure work "very painful" and noting the tool is now free and open source; the builder runs at 1,000 tokens per second, and a linked demo connects to Vercel AI Gateway's 200+ models and outputs pure HTML artifacts — via 1 2
4. Exa 3.0 launched, described as the company's largest improvement in search quality to date and ranked first across all tested benchmarks, including coding and GTM intelligence verticals, stemming from multiple research breakthroughs — via 1

2. Compute Economics and Investment

1. IREN's GPU purchases are typically backed by multi-year take-or-pay customer contracts with prepayments covering 45–55% of related GPU capex, with the remainder financed against contracted GPUs; the Microsoft contract carries roughly 6% A-grade financing that, combined with prepayments, covers 96% of GPU capex, while Mackenzie's non-investment-grade customer financing is a 9% fixed rate covering all GPU costs with prepayments. Demand reportedly exceeds build capacity, with 2026 capacity essentially sold out and three-year contract prices more than doubled year over year — via 1
2. Tiny invested $250K in Poolside in 2023; under Nvidia's financing terms, that stake carries an implied book value of about $11.1M (roughly C$15.8M), about 44x the original investment, and Tiny continues to hold it. The Nvidia deal includes a $6B technology license and a $1B investment in Poolside at a $13B post-money valuation; Tiny also holds shares in spun-out Poolside Infrastructure, which is pursuing the Project Horizon campus in West Texas — 200 MW first within a 1.2 GW site, targeting operation by end-2027, with over 7 GW of long-term expansion potential subject to financing and execution — via 1

3. Agent Products and AI Adoption

1. Wabi 2.0 was praised by Tobi Lütke as well-made and worth trying; it is positioned as a multi-person, collaborative agent that multiple people can join in the same conversation thread for syncing family schedules, booking event tickets and planning and booking travel. Garry Tan commented that multi-person collaborative agents in group-chat form are "quite inevitable" — via 1 2
2. Willow launched its writing assistant Wally, which lets users iterate in conversation until satisfied, helps organize ideas, drafts in the user's writing style and learns their habits and preferences over time; it also launched on Product Hunt — via 1
3. Ryan Hoover noted that SMB AI adoption is far lower than expected: he had assumed a consumer-agent calling and texting boom would drive SMB adoption over the next few years, but many businesses lack even basic online booking — via 1
4. Levelsio reported that a friend doing ordinary software work at a traditional company in an Austrian town went "fully AI": tasks now take 5 seconds, the employer will cancel a $10K/year CMS license in favor of vibe-coded dashboards, and by 2027 coding will be about taste, with the author speculating AI coding may have truly gone mainstream and noting he has not read code for over a year — via 1

4. Platform and Product Updates

1. Sahil Lavingia shared that sellers can now set discount codes limited to new customers via a "Limit to new customers" toggle, so first-time buyers get the discount while existing customers do not — via 1
2. Grok Bot 0.71.2 supports up to 20 files per message, files up to 250 MB and video up to 1 GB without compression or splitting, plus smarter notifications and form prefill; another update covers new chat cards, speed, draft editing, email aliases, X DMs, @bot, connectors, voice calls, notifications, invites, Android Assistant and access lists — via 1 2
3. Alertmouse's MCP connector is live, letting users connect @alertmouse in Claude or ChatGPT to use its mentions and AI features — via 1