Daily Brief Jul 11, 2026
Palantir CEO Alex Karp says enterprises fear frontier labs will train on their data; only a 20VC roundtable relays the interview
This issue was not emailed. An automated pre-send check did not pass, so subscribers did not receive it. The text is archived here unchanged.
At a glance
1. HubSpot withdrew, within a week, a plan to share customers' sales-prospecting data with other customers, according to a July 9 roundtable on 20VC, a venture-capital podcast. The same roundtable relays Karp's CNBC claim that big companies distrust Anthropic and OpenAI more than ever; Palantir sells the remedy for that distrust.
2. Microsoft is putting US$2.5B and 6,000 engineers into enterprise customers' offices, according to the same 20VC roundtable; Microsoft's own announcement hasn't been reviewed, and this issue has no public comparison for that size.
3. Anthropic's November 2025 announcement page calls Opus 4.5 the best model for coding, agents and computer use, and Google benchmarked Gemini 3 against GPT-5.1; the circulating claim that both overtook GPT-5 names the wrong OpenAI model.
This issue draws on SecondSource's own research brief for July 11, 2026. The current events come from a July 9 roundtable on the 20VC podcast and the July 10 AI Engineer conference. Items drawn from a 2024 archive backfill and a check of two 2023 papers against arXiv are each marked with their original dates. Named commentary carries a new piece this week, from swyx on July 10. This issue uses 15 outside sources, each with a link to check. The full edition of this issue is the archive of record.
SecondSource isn't a news digest: each day it hunts the AI firehose for the insights that matter and the practitioner judgments worth tracking over time, and shows how every item was verified. The point is always which judgment got harder and who's been right, not what happened today.
Today's main line
1. [This week] (roundtable of July 9) HubSpot's reversal the same week gives a real trigger to the data fears Palantir's Karp described on CNBC
Why it matters: Data and IP trust joins model reliability and process change as a barrier to enterprise AI, pushing demand toward data kept in-house.
Read the full item
A July 9 roundtable on 20VC, the venture-capital podcast hosted by Harry Stebbings (video), relayed Palantir CEO Alex Karp's widely shared CNBC interview. As the roundtable tells it, Karp said large enterprises' skepticism toward frontier model makers, naming Anthropic and OpenAI, is the highest it has ever been. The skepticism has two parts: what the spending actually buys, and whether the labs take the data customers hand over, train on it, learn the customer's business and sell that to everyone else.
In the same week, HubSpot tried to pool customers' sales-prospecting data and share it with other customers. Customers revolted, and HubSpot withdrew the plan within a week (relayed in the same roundtable).
Not proven yet: Karp's remarks reach this item only through one roundtable's spoken relay
Verification: Karp is an interested party. Palantir sells exactly the remedy for distrust of these vendors, and Palantir's stock rose 9% that day, per the roundtable; that figure is spoken and single-source. Karp's words deserve a discount. The HubSpot withdrawal is a separate event, so the fear has a real trigger. The roundtable also accused model makers of lying about where their training data comes from. That charge rests on this one source, so read it for direction only.
Judgment update: SecondSource had tracked two axes of the barriers to enterprise AI adoption: model reliability and internal process change. Today it adds a third, data and IP trust. Our read: that axis will push demand toward keeping data in the customer's own hands and away from lock-in to a single vendor.
2. [This week] (roundtable of July 9) Microsoft is reported to be embedding engineers with customers, a way of selling the roundtable likens to IBM's services arm of 30 years ago
Why it matters: If the roundtable's account holds, selling AI to enterprises needs a services layer, and that layer earns far lower margins than models.
Read the full item
The same roundtable (video, 2026-07-09) relayed that Microsoft has announced a US$2.5B program that places 6,000 engineers on site with enterprise customers. The roundtable presents the program as an open answer to an MIT finding that 95% of enterprise AI pilots show no measurable impact on profit and loss. Amazon moved in the same direction two days earlier.
The roundtable's reading: OpenAI and Anthropic are product companies to the bone, so enterprise adoption needs a services layer in between. Microsoft is stepping into the role that IBM's and HP's services divisions once played. In the roundtable's words, every tech company either goes under or lives long enough to become the next IBM. The cost is margins far below those of selling models.
Not proven yet: US$2.5B, 6,000 engineers and MIT's 95% are one roundtable's spoken relay
Verification: All three figures come from venture capitalists speaking on one roundtable, so treat them as reference only. Microsoft's own announcement hasn't been reviewed. A follow-up update will run once that primary document is in hand.
Judgment update: Where enterprise adoption gets stuck is now a three-way collision. The roundtable bets on process change plus a shortage of on-site talent. Against it stands Sarah Guo's view, already on SecondSource's long-term watchlist from late 2024 (Colossus interview): the bottleneck is model reliability, not process. Item 1's data-trust axis is the third. Our read: the three more likely coexist, each dominant in different settings. Which layer dominates decides who captures the money: the services layer, the evaluation layer or the private-deployment layer. The verdict points are the on-site program's renewal rate after a year and the slope of Anthropic's revenue ramp.
3. [This week] (July 9–10) One roundtable produced three sets of company rumors in a day, and SecondSource takes none of them as fact
Why it matters: Custom chips, NVIDIA's revenue recognition and a US stake in OpenAI all rest on one spoken relay, so none feeds any judgment here.
Read the full item
All three come from the July 9 and 10 material, from the same 20VC roundtable.
(a) DeepSeek has started designing its own chips, and Anthropic is in talks with Samsung about custom chips.
(b) NVIDIA uses a use-compute-now, pay-later style of revenue recognition, and the share of revenue from NVIDIA's top three customers fell from 80% to 50%.
(c) OpenAI would let the US government hold a 5% stake. The person relaying this conceded not having read the original source.
All three are spoken relays from a single roundtable. None enters any judgment in this newsletter until checked.
Not proven yet: all three rest on one roundtable's spoken relay, with no second source
Verification: The item itself is the verification method. Claim (b)'s customer concentration can be checked directly against NVIDIA's quarterly filings with the US Securities and Exchange Commission. The other two wait for a second independent source. Wherever these three claims circulate, their status is the same: a single relay, unverified.
4. [Evidence update] (event of December 2, 2025) In OpenAI's code-red episode, the model rivals were measured against was GPT-5.1, not the GPT-5 of the circulating account
Why it matters: The event holds across outlets, but which OpenAI model was overtaken slipped one step in retelling, a drift only the official originals catch.
Read the full item
The event itself: on December 2, 2025, Sam Altman's internal memo warned of competition and paused ads and a shopping assistant so OpenAI could focus on ChatGPT. Fortune, Forbes and other US outlets reported it the same day, and SecondSource confirmed it from multiple sources in early July.
The change today is a precision fix on wording. The circulating version says Gemini 3 and Claude overtook GPT-5. Checked against the originals, Gemini 3 (November 18, 2025) was benchmarked against GPT-5.1. Anthropic's own announcement only calls Opus 4.5 (November 24, 2025) the best model in the world for coding, agents and computer use (official page, 2025-11). The announcement makes no claim about GPT-5.
Not proven yet: the performance-per-watt and NVIDIA-moat inference has no primary source
Verification: The event itself rests on multiple outlets. Who overtook whom slipped one level along the retelling chain, which is why this newsletter goes back to official originals. The accompanying inference that the battleground is moving to performance per watt and that NVIDIA's moat is wobbling remains analysts' interpretation, with no primary source to check it against. It is neither adopted nor rejected here.
5. [Trend watch] (2024 archive) Three 2024 answers on where AI money should go now join SecondSource's tracked judgments
Why it matters: Three 2024 practitioner answers on retrieval, code and tool interfaces held up by 2026; the two speakers behind the first sell such systems.
Read the full item
Practitioners who didn't know one another gave these answers separately in 2024, and looking back from 2026 all three have held. The cross-source check was completed today.
(1) Retrieval or fine-tuning is a false debate. Retrieval (RAG) governs what a model knows: it pins facts into the context at question time. Fine-tuning governs how a model behaves, the way a stem cell differentiates into a specific cell type. Fine-tuning changes behavior patterns; knowledge goes in through retrieval. Mark Huang of Gradient (2024-05-30) and Mike Conover of Brightwave (2024-06-11) laid out the same framework independently in the same period.
(2) Code is the general interface for agents. Wrapping the model in code buys reliability: the workflow lives in code and the model handles only the semantics. Letting the model write code buys room to act: the model can build the tool it needs instead of picking from a list of 128. Sam Schillace of Microsoft made the first point (2024-03). Imbue and Databricks made the second at the same event (2024-06).
(3) The tool interface is about 90% of an agent's success. Shunyu Yao, author of ReAct and SWE-agent and then at OpenAI, put it this way: "making the tool good and reliable is probably like 90% of the whole agent" (2024-09-27). The same conversation credits Devin's breakthrough to the interface rather than to planning.
Not proven yet: the two speakers behind answer (1) both sell retrieval and fine-tuning systems
Verification: Six independent origins, none citing another, converged in the same year. Answer (1) carries a stance discount, because both speakers sell this kind of system. In hindsight, agents that have the model write and run code are now mainstream, and fine-tuning to inject knowledge is widely regarded as an anti-pattern.
Judgment update: The three answers work as a buyer's checklist. Evaluating an agent product turns on where its behavioral boundaries sit, how large its code-execution surface is and how that surface is secured, and who built its tool interface. A training budget turns on whether the goal is changed behavior or injected knowledge.
Named commentary
swyx (July 10, 2026, AI Engineer conference): "I genuinely do expect Fable to be the end of this era of [LLMs]." swyx, who organizes the AI Engineer conference, made the argument on site (video, 2026-07-10). In swyx's account, what ends the keep-getting-bigger path is not a capability ceiling but the economics of usability. The smartest models are slow and expensive, and time is the hard budget everyone shares: someone able to pay unlimited money still can't afford unlimited waiting. The way forward, swyx says, runs through data efficiency, continual learning and new architectures, and none of them is in hand yet. On data efficiency, humans learn from text on the order of millions of words, while models train on trillions of tokens. swyx also cooled talk of AI self-improvement setting off an explosion: most current self-improvement recombines things near known territory.
Which of our calls this connects to: swyx's view points the same way as Ilya Sutskever's recent argument that the generalization gap, not scale, is what holds models back. Within one week, two independent insiders each said the scaling era is giving way. One discount applies: this is a single expert's prediction, so read it for direction.
Model watch
1. [Evidence update] (papers from 2023) Two widely cited, rarely read 2023 papers, checked word for word against arXiv: they are about two different things. These two papers are still cited as evidence in 2026, and the claims drift as they are passed along. Going back to the originals recalibrates current judgments; it isn't archaeology.
*Part of the foundation for how reasoning is trained: Let's Verify Step by Step*** (OpenAI, 2023-05, arXiv). The paper shows that rewarding every correct step of reasoning during training (process supervision) teaches a model to solve hard problems better than rewarding only the correct final answer. On a representative subset of the MATH benchmark the method reached 78%, and the authors released 800,000 step-level human labels. This is a training-side method, not an inference-side one. It is one piece of the foundation for the later o1 approach of spending training compute to teach models to think, but OpenAI hasn't fully disclosed the reward recipe o1 actually uses (see the research trends below).
*Don't trust headlines that a new architecture crushes Transformers: Never Train from Scratch (ICLR 2024, 2023-10, arXiv). This paper has nothing to do with reasoning.* It is about how to compare architectures fairly. Many papers claim that a new architecture, such as state-space models in the Mamba family, beats Transformers, but in those comparisons both sides are trained from scratch. This paper shows that once both sides get a proper round of pretraining first, the new architecture's advantage shrinks sharply or disappears. Part of the impression that new architectures win big is therefore an artifact of an experimental setup that never let the Transformer warm up. The question for any 2026 claim that a new architecture beats Transformers across the board is whether pretraining was controlled for.
Verification: Both papers were checked word for word against their arXiv originals today, and the figures (78%, 800,000 labels) come from the paper itself.
Research trends: SecondSource maintains a map of AI research across 13 fields and 59 method directions, each marked mainstream, rising or fading and written in terms of consequences. Three with the most commercial weight, each with a date and its evidence:
2. [Trend] Training: rewarding whether an answer is right, instead of imitating human demonstrations, is the core recipe behind this wave of reasoning models solving hard math and competitive programming. The timing matters. This isn't a 2023 development: it became mainstream after o1 led in September 2024 and DeepSeek R1 followed in January 2025. The evidence for this recipe: the R1 paper (2025-01) states that R1 uses rule-based rewards, judging math by whether the answer is right and code by whether it compiles, with no human preference labels. For o1, OpenAI has said only that it trains reasoning with reinforcement learning; the exact reward recipe isn't disclosed, and the details are outside reverse-engineering. So for this trend, R1 is direct evidence and o1 is a reasonable inference. [Business] This is the technical source of the whole reasoning-model product category, o1, R1 and what followed.
3. [Trend] Customization: to make a model use a company's own knowledge, the mainstream approach has moved from retraining the model to feeding data in at question time (retrieval) plus letting the model call tools. Practitioners reached that consensus in 2024 (see the six independent origins in main-line item 5), and by 2026 it is the default. [Business] It maps directly onto enterprise knowledge-base Q&A and agent products: for making AI understand a company's internal material, fine-tuning on that data is now almost nobody's first idea.
4. [Trend] Chips: as models grow, the bottleneck moves from how fast chips compute to how fast data reaches them, and high-bandwidth memory (HBM) has become the tightest component. Compute units often sit idle waiting for data to arrive from memory, so whoever feeds data faster wins. Only a few companies in the world, SK hynix, Samsung and Micron, can make HBM. [Business] The cost and supply of compute increasingly depend on these few memory makers, which is why AI training costs rise when memory prices do.
Chips & semiconductors
1. [Trend watch] (October 2024) "Don't buy H100s": cheaper compute and pricier power are two sides of one coin. Eugene Cheah, co-founder of the inference provider Featherless, gave a first-hand snapshot of the GPU rental market (guest post, 2024-10). H100 hourly rents fell from US$8 during the shortage to US$2, a drop of at least 40% a year. Cheah wrote: "NVIDIA's marketing projection of $4 per GPU hour across 4 years, has evaporated away in under 1.5 years." And: "GPUs are commodity hardware. Faster than all of us expected." The same month, semiconductor analyst Doug O'Laughlin recorded the mirror image on the power side (Fabricated Knowledge, 2024-10). The Three Mile Island nuclear plant is restarting for AI, and capacity auction prices on PJM, the grid operator for the eastern US, rose roughly tenfold in a year, from US$28.92 to US$269 per MW-day. O'Laughlin's key call: nuclear's real bottleneck is financing, not technology, and the record of Vogtle, whose cost climbed from US$14B to US$35B, keeps capital away.
Verification: Each piece is a single author's first-hand account, built on personal experience and the author's own arithmetic, so discount the figures. The two are independent of each other and point to the same structure in the same month. Looking back from 2026, H100 commoditization was called correctly, but compute capital spending didn't shrink: it moved up to clusters of the newest chips. Scarcity moved from silicon to the grid and to capital.
What to watch: two prices. Hourly rents for older GPU generations falling signals commoditization; grid capacity auction prices rising signals the new bottleneck.
Product moves
No product news this issue.
From the archive
1. Capability is no longer a fixed property of a model but a function of how much compute is spent, and that rewrites evaluation, safety and loss of control at once. Noam Brown, who leads reasoning research at OpenAI, set out the claim on No Priors (No Priors, 2026-06-26). With the right scaffolding, a model can keep thinking about a single problem for weeks without stalling. The same model given US$10 of compute and given US$100,000 of compute differs enormously in capability. Three consequences follow. First, saying model A beats model B without controlling the compute budget is a bad comparison, and single-score leaderboards lose their meaning. Second, a safety evaluation that tests dangerous capabilities at a fixed budget reports false safety when those capabilities keep rising with budget. Third, there will be no overnight loss of control, because the strongest capabilities cost real time. The measurement anchor is METR's longitudinal study (arXiv, 2025-03): the length of tasks models can complete autonomously doubles roughly every seven months, but at publication the ceiling at a 50% success rate was only about 50 minutes. The direction holds; week-long autonomy is still far off.
Verification: Dozens of 2025 reasoning papers cut both ways. Some support the claim, with capability rising with budget across fields; others set limits, with some fields getting worse when models think longer. The core holds; unlimited extrapolation doesn't.
What it means: A benchmark score from a model launch says little without the compute spent to reach it. Articles predicting an overnight AI takeover leave out the clock: the strongest capabilities take real time to produce.
Sources & accounting
The past 24 hours. The routine fetch had zero failures. Overnight, 32 pieces and episodes were read in full: podcast roundtables, newsletters, X and blogs, counted by piece or episode and not compared across types. During the day, 5 more pieces from the 2024 archive were read, for 37 in all. Only 3 current events and 3 sets of rumors were strong enough for the body; material with nothing new was filtered out. Retrospective material is not counted in these figures.
The sources we track. The judgments in this newsletter rest on 529 named voices and sources on SecondSource's long-term watchlist: 304 people on X, 90 podcast voices, 48 paper authors, 48 blogs, 46 newsletters, 26 companies' earnings reports and calls, and others. No sources were added this issue. This issue uses 15 outside sources in the body.
— SecondSource · generated by our research system · 15 sources · Replies go straight to the editors as feedback
Written from the same research and judgments as the Traditional Chinese edition. Sources are linked; we distinguish original documents from reporting and mark what we could not verify.