We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
1. Funding & Deals
The period’s clearest deal signal is a reported Stripe acquisition of OpenRouter for more than $7B—but it is a strategic datapoint, not an early-stage financing comp. A current-period venture post relaying Bloomberg says the deal followed a $1.3B round roughly 82 days earlier; it describes OpenRouter as founded in 2023, with 8M users, more than $100M in annualized inference volume, and backing from Sequoia and a16z. Its product is a routing layer: one API in front of 400-plus models, rather than a model lab. The post’s thesis is that aggregation, routing, price discovery, and settlement may capture durable AI value as models commoditize.
Do not use the reported 5x-in-82-days multiple for seed underwriting. The useful question is whether this represents genuine repricing of the routing layer or a strategic premium from Stripe; the source leaves that unresolved.
2. Emerging Teams
An AI-native Canadian law firm combines unusually strong founder–problem fit with an early, self-reported traction signal. Its founder has 25 years of contracting experience, including a decade as general counsel at a large BC private company and a prior CEO role in international aviation. The firm uses flat fees and a 48-hour target, has AI perform the first pass from its own playbooks, versions every draft with an audit trail, keeps client data out of model training, and retains founder review and signature. Demand was described as constant, with the first month tracking toward low-to-mid five figures.
This is a service-delivery model rather than legal-copilot SaaS: the founder says the platform was built for the firm’s own use, not to sell to other law firms, with startups and SMBs as the target market. The diligence question is whether the flat-fee, rapid-turnaround workflow remains reliable as volume rises without weakening legal review.
Factory Brain is a small but clear infrastructure thesis around “learn once, reuse later.” The private-preview project stores business terminology, schema relationships, KPI definitions, validated questions and SQL, corrections, and follow-up context so that an LLM is reserved for new or genuinely complex questions. It is also being designed around RLS/CLS, governance, and keeping business data in the customer environment. The founder is explicitly asking for architecture criticism rather than presenting traction; the current product signal is the attempt to make semantic memory and query reuse part of the analytics stack, not another chat-with-data wrapper.
3. AI & Tech Breakthroughs
Document agents are being evaluated against completeness and evidence, not just plausible answers. LlamaIndex’s new ExtractBench is an open benchmark covering 370 enterprise documents, 4,869 pages, eight business domains, 67 document types, and 14 systems; its ground truth combines cross-model agreement with human adjudication, synthetic long lists, and manually checked forms. The vendor reports that Agentic Plus leads at 95.6% value F1, with the best grounding scores at 8.1¢ per page.
The more important result for diligence is the failure profile: LlamaIndex reports that commercial VLMs fall below 35% recall on documents longer than 50 pages, while coding agents return no evidence by default; it says Agentic Plus holds 94.4% on the longest documents. Those are vendor-reported results, but the dataset, harness, and paper are public, so long-document recall and grounding can be independently tested rather than accepted from a demo.
Long-context efficiency still has an exact-retrieval problem in genomics. A developer’s 1M-token DNA experiment reports roughly 25% needle-in-a-haystack recall—chance level for a four-token DNA vocabulary—and similarly poor 25–27% results for HyenaDNA, versus 50–60% recall for a much shorter 16K context. The accompanying discussion points toward selective state-space models or hybrid architectures that retain occasional uncompressed attention. This is a self-reported research thread, not a validated benchmark, but it is a useful warning against treating million-token context as equivalent to million-token memory.
Text provenance also remains brittle. A developer reports that, across nearly 300 watermark tests, inserting invisible Unicode variation selectors into about 30% of characters reduced a watermark score from 45 to below 1 in every one of ten trials; the same post says code is often barely watermarked because its token distribution is low-entropy. Treat this as an attack report requiring replication, not as a general defeat of all watermarking schemes.
4. Market Signals
The inference-glut question is becoming a two-market asset-quality problem. An Investing in AI analysis expects AI data-center power to remain tight through 2027 and loosen by mid-2028 without an aggregate glut, while warning that the average hides a split between new facilities able to host 130–600 kW liquid-cooled racks and older capacity. It estimates 120 GW of scheduled capacity through August 2028, but only 50–60% realization because of transformers, turbines, and interconnection queues; the resulting AI fleet would rise from 35.6 GW to 78.5 GW into a market described as 94% full.
Demand is not automatically falling with inference prices: the essay says token prices fell from about $20 to $0.07 per million while Google’s reported monthly volume rose 330-fold, with measured elasticity of −1.03. Its model also estimates that raising reasoning queries from 12% to 45% lifts average energy per query 2.6x. The positioning implication is to underwrite workload mix and routing, not simply installed megawatts.
The analysis projects roughly 4 GW of older provisioned capacity becoming effectively stranded, with price pressure beginning around Q2 2027—well before aggregate vacancy shows weakness. Its recommended monitors are hyperscaler capex split between training and serving, H100-class spot GPU-hour pricing, preleasing on capacity under construction, and interconnection or turbine-order cancellations. The main caveats are that the realization haircut is a judgment call and token growth leans heavily on unaudited Google figures.
AI adoption is asymmetric by workflow. Andrew Chen’s framing is that workplace AI succeeds where it compresses patterned drudgery—forms, process steps, boilerplate, and updates—while consumer products need novelty, parasocial connection, and authenticity. He extends the same problem to sales and marketing: uniform AI messaging is easy to generate but fails in adversarial settings where the message must be fresh and differentiated.
The “agent test” is becoming a product and valuation filter. Jason Lemkin says SaaStr’s agents built an ad-creative operation without Canva or Notion ever entering the workflow, despite SaaStr having paid for and liked both products. He argues that no-code products built around replacing a missing specialist are exposed to native AI substitution, while Gartner data suggests fewer than 10% of enterprises have successfully deployed an agentic application. His practical test is to give an agent the job a product does without instructing it to use that product, then mark valuations from current growth rather than legacy financing marks.
5. Worth Your Time
- Watch — Lenny’s Podcast: OpenAI’s Head of Product Design, Ian Silber. Silber describes a future ChatGPT as a proactive, voice-rich universal input that decides whether to answer or act, hides model and mode choices from most users, and supports durable repeatable workflows rather than one-off chats.
Read — ExtractBench. Use it as a concrete evaluation starting point for long-document completeness, grounding, perception failures, and cost—not just clean-PDF accuracy.
Read — Is There An Inference Glut Coming?. The useful parts are the distinction between modern and stranded capacity and the monitorables that could reveal softness before aggregate utilization does.
