Parameter — Daily AI Brief
Today in AI
- FrontierMath v2 — Epoch re-benchmarks after fixing 42% of problems; Fable 5 leads at 88% (Tier 4).
- Kimi K2.7 Code — Moonshot's 1T-param open-weight MoE beats Opus 4.8 on MCPMark, cuts reasoning tokens ~30%.
- DeepSeek mHC — manifold-constrained hyper-connections fix deep-residual instability at ~7% fixed overhead.
- Oracle Q4 FY26 — record $19.2B revenue and a $638B backlog, but the stock falls 10% on capex.
- Neura Robotics — up to $1.4B Series C from NVIDIA, Amazon, Qualcomm, Tether at a ~$7B valuation.
- Anthropic IPO — confidential S-1 at a $965B valuation; revenue annualizing ~$47B.
- Gemini 3.5 Pro — 2M-token context and Deep Think enter the final GA countdown.
Tape. Oracle fell ~10% after-hours June 10 despite a record quarter — a $70B FY27 capex guide and a $40B equity raise turned the AI buildout's financing into the day's dominant question [16][18].
1. Epoch rebuilds FrontierMath as "v2" — and the ceiling jumps again
What happened. On June 12, Epoch AI released FrontierMath v2, a corrected edition of its expert-authored math benchmark after an audit found "small but critical" errors in 42% of the original problems; the cleaned set now totals 338 problems across Tiers 1–4 [1][2]. On the refreshed leaderboard, Anthropic's Claude Fable 5 takes the top spot at 87% on Tiers 1–3 and 88% on Tier 4 [2].
The technical read. FrontierMath is deliberately adversarial to memorization: original, unpublished problems spanning number theory, real analysis, algebraic geometry and category theory, each requiring hours-to-days of expert effort, graded against private answer sets to resist contamination [1]. Tier 4 is the "research-grade" band that stood near single digits a year ago — GPT-5.5 Pro was cited leading the original Tier 4 at 39.6% earlier this spring [3]. An 88% v2 Tier 4 result is therefore a step-change, though the audit itself is a confounder: Epoch notes rankings held but "scores are higher across the board" after error removal, so part of the jump is measurement, not capability [1][2].
Why it matters. When 42% of a flagship benchmark's items needed correction, every prior FrontierMath number in circulation carries a quiet asterisk — model cards, marketing decks and procurement scorecards alike [1]. Benchmarks are now infrastructure, and their maintenance burden is becoming a first-class research cost.
Parameter View: The headline is not "Fable 5 hit 88%" — it is that the world's hardest public math benchmark needed a 42% repair in under a year of heavy use. Saturation pressure is forcing eval turnover faster than labs can re-baseline, and a model that "leads" on v2 may be partly cashing in the ~5–10 points the cleanup added rather than new reasoning (Parameter estimate, from Epoch's "higher across the board" note). Buyers should demand the edition with every FrontierMath claim; a score without "v1/v2" is now uninterpretable.
2. Moonshot ships Kimi K2.7 Code — a 1T-parameter open model that out-tools Opus on agents
What happened. Moonshot AI released Kimi K2.7 Code on Hugging Face on June 12 — a coding-and-agents-focused refresh of its trillion-parameter open-weight line, under a Modified MIT license permitting commercial use with attribution [4][5].
The technical read. K2.7 Code keeps the 1T-parameter Mixture-of-Experts backbone with ~32B active parameters per token and a 256K context window, and ships with native INT4 quantization for serving on vLLM, SGLang or KTransformers [4][5]. Its standout result is MCPMark Verified — a tool-use suite run against live Notion, GitHub, filesystem, Postgres and Playwright servers — where it scores 81.1 and beats Anthropic's Claude Opus 4.8 at 76.4 [4][5]. On MLS Bench Lite it posts 35.1, within a point of GPT-5.5's 35.5, while cutting reasoning-token usage roughly 30% versus the prior K2.6 generation — a direct hit to inference cost on long agentic traces [4][5]. The weights are large (~340GB at the shipped quant), so "open" here means lab- and enterprise-deployable, not laptop-friendly [6].
Why it matters. An open-weight model that wins a real-environment tool-use benchmark against a frontier closed model — and does it ~30% cheaper per trace — narrows the gap exactly where agents are being commercialized: multi-step tasks against real services, not chat [4][5].
Parameter View: The competitive front has moved from "who writes the best code" to "who completes the longest task chain per dollar," and that axis structurally favors open weights, because token-efficiency compounds with self-hosting. A 30% reasoning-token cut on a 32B-active MoE is worth more to an agent operator than two points of pass@1: at scale it is roughly a 30% line-item cut on the dominant cost of agentic inference (Parameter estimate). Expect closed labs to answer with token-budget controls, not just bigger scores.
3. DeepSeek's manifold-constrained hyper-connections fix the instability that broke deep residuals
What happened. DeepSeek's mHC paper — Manifold-Constrained Hyper-Connections, co-authored by founder Liang Wenfeng — has become the most-discussed scaling-architecture result of the year, with analysts framing it as a path to training bigger models for less [7][8][9]. It rethinks the residual connection, the load-bearing primitive of every modern Transformer.
The technical read. Plain residuals add a layer's output back to its input through one stream. Hyper-connections generalize this to multiple parallel streams with learned mixing — more expressive, but unstable: in a 27B-parameter model, unconstrained hyper-connections produced signal gains exceeding 3000x, causing catastrophic divergence [7][9]. mHC constrains the mixing to a manifold that redistributes information across streams while holding the total signal energy constant, so depth no longer amplifies [7]. Tested at 3B, 9B and 27B parameters, it trains smoothly where the unconstrained variant blows up, and the compute overhead is small and roughly constant with scale — about 6.7%, i.e. a 1,000 H100-hour job becomes ~1,067 H100-hours [7][9].
Why it matters. The expensive failure mode in frontier pre-training is the run that diverges after weeks of GPU time. A primitive that keeps deep, wide networks numerically stable for a fixed ~7% tax attacks training cost variance, not just the loss curve — and it is architecture-agnostic enough to drop into other labs' stacks [8][9].
Parameter View: The quiet tell is that the overhead is constant across 3B→27B. If that flatness holds to the hundred-billion-parameter range, mHC is less a tweak than a depth-scaling unlock — the residual-stream analog of what normalization did for trainability a decade ago. Watch for Western labs to quietly adopt manifold-style constraints in their next pre-training run; "training bigger for less" is a margin story before it is a capability story, and DeepSeek just published the recipe.
4. Oracle prints a record quarter and a $638B backlog — and falls 10%
What happened. Oracle reported fiscal Q4 2026 on June 10: record total revenue of $19.2B (up 21%), with adjusted EPS of $2.03 against a $1.96 estimate [16][17]. The stock fell ~10% after hours [16].
The technical read. The demand signal was extraordinary — remaining performance obligations (RPO, the contracted backlog) hit $638B, up 363% year over year and $85B sequentially, and OCI (cloud infrastructure) revenue reached $5.8B, up 93% [16][17]. The market sold the financing, not the demand: Oracle guided FY2027 capex to ~$70B, said it would raise ~$40B in equity — double the prior ~$20B expectation — and disclosed FY2026 free cash flow of negative $23.7B as the AI buildout outran cash generation [16]. Q1 FY27 EPS guidance of $1.72–$1.76 sat below the $1.80–$1.90 consensus, with cloud costs up 56% against cloud revenue up 47% — i.e. near-term margin compression [16].
Why it matters. Oracle is the cleanest public proxy for "neocloud" AI infrastructure economics. A 363% backlog with negative free cash flow and a doubling equity raise is the whole AI-capex thesis in one print: the orders are real, but so is the bill, and who funds it is now the swing variable [16][17].
Parameter View: A $638B backlog against -$23.7B free cash flow means Oracle is effectively pre-selling compute it must borrow and dilute to build — a leveraged bet that GPU depreciation schedules stay behind contract revenue. The selloff is the market repricing duration risk, not demand. If even backlog-rich names must raise equity into the buildout, the marginal AI dollar is getting more expensive — bullish for the chip suppliers paid up front, more cautious on the buildout middlemen carrying the balance-sheet risk.
5. Neura Robotics raises up to $1.4B — and NVIDIA, Amazon and Tether all show up
What happened. German humanoid-robotics firm Neura Robotics closed a Series C of up to $1.4B on June 10, at a valuation of roughly $7B, with NVIDIA, Amazon, Qualcomm, Tether, Bosch, Schaeffler and the European Investment Bank participating [10][11].
The technical read. The round sits inside a record year for the category: robotics startups have raised more than $23B in 2026, already approaching the full-year 2025 total, with humanoids drawing the bulk [11]. The investor list is the signal — NVIDIA (compute and the physical-AI stack), Amazon (logistics deployment), Qualcomm (edge inference silicon) and two German industrial primes (Bosch, Schaeffler) — pairing a frontier-AI supply chain with concrete industrial demand [10]. It lands against Figure AI's reported ~$39B valuation, marking how steeply embodied-AI multiples have re-rated [11].
Why it matters. Capital is flowing to the layer where AI meets the physical world, and it is increasingly strategic rather than purely financial — chipmakers and industrials buying optionality on humanoid labor [10][11].
Parameter View: When a company's GPU vendor, its cloud-logistics customer and its edge-silicon supplier all sit on one cap table, the round is a supply-chain alliance dressed as equity. Humanoids are now a way for NVIDIA and Qualcomm to seed downstream demand for inference silicon at the edge — the economic logic that made data-center AI compounding, exported to the factory floor. The risk is a 2021-style funding peak: $23B into a largely pre-revenue category is a bet on deployment timelines that have slipped before.
6. Anthropic's confidential S-1 lands at a $965B valuation with revenue annualizing near $47B
What happened. Anthropic confidentially filed for an IPO on June 1, following a $65B Series H that set its valuation at $965B; bankers expect a public debut above $1T, with an October 2026 window widely discussed [12][13].
The technical read. Independent estimates put Anthropic's annualized revenue at roughly $47B as of May 2026, up from about $9B at the end of 2025 — a ~5x run-rate expansion in five months — with the company telling investors it expects to cross a $50B annualized rate and report its first profitable quarter in this June quarter [13][14]. Total capital raised to date is on the order of $125B, much of it earmarked for compute [13]. The operating structure is distinctive: Anthropic buys compute from suppliers rather than owning data centers, the inverse of OpenAI's proprietary-infrastructure approach [14].
Why it matters. A near-$1T private valuation on ~$47B annualized revenue — roughly 20x run-rate revenue — is the public market's first true referendum on frontier-lab economics, and the compute-light operating model will be scrutinized line by line in the prospectus [12][13].
Parameter View: The "buy, don't build" compute strategy is the most important disclosure in the filing. It converts capex into opex and keeps the balance sheet light — the mirror image of Oracle's leveraged buildout (Item 4) — but it hands pricing power to suppliers and caps gross margin at whatever the cloud and chip vendors allow. Anthropic's path to a $1T listing runs straight through NVIDIA's and its cloud partners' price lists; the IPO is, in part, a bet that frontier-model demand outruns compute-cost inflation. First profitable quarter or not, that is the spread the market will price.
7. Gemini 3.5 Pro enters its GA countdown — 2M-token context and "Deep Think"
What happened. Google's Gemini 3.5 Pro, unveiled at I/O on May 19, is in the final stretch toward general availability, with reporting pointing to a launch before June 30 after an extended Vertex preview [15][19].
The technical read. Pro targets a 2-million-token context window — the largest of any production frontier model — plus a "Deep Think" extended-reasoning mode and frontier multimodal understanding across text, images and other formats [15]. Google has been deliberately tight-lipped: no public model card, API pricing or official benchmark suite as of mid-June, though the earlier Gemini 3.5 Flash already beat Gemini 3.1 Pro on coding and agentic tasks while regressing on hard reasoning — precisely the gap Pro is meant to close [15][19]. Expected pricing has been floated around $15/$60 per million input/output tokens, well above the prior Pro tier [19].
Why it matters. A 2M-token, Deep-Think Pro tier is Google's answer to the reasoning-and-context arms race headlined by GPT-5.5 and Claude Fable 5; with Gemini increasingly the default model behind Search and Workspace, GA is a distribution event as much as a capability one [15][19].
Parameter View: The signal to watch on launch day is not the top-line benchmark — it is whether 2M context comes with usable latency and a price that does not tax it into irrelevance. A long window that costs $60/M output and answers slowly is a spec-sheet trophy; one that stays cheap and fast at length is a genuine moat for document- and codebase-scale agents. Google's silence on pricing this close to GA suggests the cost-of-context problem is still being negotiated internally (Parameter estimate).
Market Movers
The dominant move was Oracle's ~10% after-hours drop on June 10: a record quarter and a $638B backlog were overwhelmed by a doubling of FY27 capex to ~$70B, a planned ~$40B equity raise and -$23.7B FY26 free cash flow — the market repricing how the AI buildout gets financed [16][18]. It extended a fragile fortnight that began June 3–4, when Broadcom's refusal to raise its AI-chip outlook dragged the whole semiconductor complex lower before a partial recovery the following week [20][21].
| Name (Ticker) | Move | Driver — why it moved |
|---|---|---|
| Oracle (ORCL) | Down ~10% (Jun 10, after-hours) | Record Q4 (rev $19.2B, RPO $638B +363%) overshadowed by ~$70B FY27 capex guide, ~$40B equity raise (2x prior) and -$23.7B FY26 FCF; soft Q1 EPS guide [16][18] |
| Broadcom (AVGO) | Down ~13% (Jun 4) | Posted record revenue but declined to raise its 2026 AI-chip outlook; biggest drop since early 2025 [20][21] |
| Philadelphia Semi Index (SOX) | Down ~2% close, -6.3% intraday (Jun 4) | Sector-wide reaction to Broadcom's guidance before a partial rebound [20] |
| Alphabet (GOOGL) | Up (early Jun) | Rallied on an ~$85B capacity/expansion commitment as Google leaned into AI infrastructure and coding tools [21] |
| Semis (NVDA, AMD et al.) | Up (week of Jun 8) | Nasdaq Composite recovered ~0.9% June 8 as chip names reclaimed Friday's losses — read as a technical correction, not a demand reassessment [21] |
Key metrics. SOX finished June 4 down ~2% after falling as much as 6.3% intraday, its worst session since early 2025, led by Broadcom's ~13% drop [20]. On the private side, Anthropic's $965B Series H valuation and Neura Robotics' ~$7B round mark fresh milestones, with robotics startups passing $23B raised in 2026 [11][12]. Oracle's record $638B RPO is among the largest contracted backlogs ever disclosed by an enterprise-software vendor — set against negative $23.7B free cash flow [16].
Positioning
| Company (Ticker) | Read | Conviction | Horizon | Thesis (one line) |
|---|---|---|---|---|
| NVIDIA (NVDA) | Add | High | 6–18 mo | Paid up front across every storyline today — Oracle/Anthropic compute, Neura's cap table, the chip-tape rebound; the supplier captures margin buyers must finance [16][10]. |
| Oracle (ORCL) | Hold | Medium | 6–18 mo | Backlog is real ($638B), but -$23.7B FCF and a $40B raise put balance-sheet risk ahead of the demand story until capex/cash inflect [16][18]. |
| Alphabet (GOOGL) | Add | Medium | 3–12 mo | Gemini 3.5 Pro GA + ~$85B capacity bet pairs frontier capability with unmatched distribution via Search/Workspace [15][21]. |
| Broadcom (AVGO) | Hold | Medium | 3–9 mo | Record revenue but a guidance air-pocket; custom-silicon demand intact, sentiment reset until the AI outlook is raised [20][21]. |
| Anthropic (private) | Watch | Medium | 6–12 mo | ~$47B annualized and a possible first profit quarter, but a ~20x run-rate $965B mark and compute-light margins face the IPO's first real test [12][13]. |
| Moonshot / open-weights (private) | Watch | Low | 6–18 mo | K2.7 Code's tool-use win + ~30% token cut pressures closed-model pricing; the structural threat is to API margins, not capability leadership [4][5]. |
References
- Epoch AI — FrontierMath benchmark (Tiers 1–4) and methodology. https://epoch.ai/frontiermath/the-benchmark
- Epoch AI — FrontierMath Tier 4 (v2) leaderboard; The Epoch Brief, June 12, 2026. https://epoch.ai/benchmarks/frontiermath-tier-4-v2 ; https://epochai.substack.com/p/the-epoch-brief-june-12-2026
- LLM-Stats — June 2026 model landscape and FrontierMath Tier 4 standings. https://llm-stats.com/llm-updates
- Codersera — "Kimi K2.7 Code: The Complete Guide — Benchmarks, Pricing & How to Use (2026)." https://codersera.com/blog/kimi-k2-7-complete-guide-2026/
- Kingy AI — "Kimi K2.7 Code Released: Benchmarks, Specs, and How It Compares." https://kingy.ai/ai/kimi-k2-7-code-benchmarks-specs/
- Modem Guides — "Kimi K2.7-Code: Open Weights, 340GB Reality Check." https://www.modemguides.com/blogs/ai-news/kimi-k2-7-code-open-source-release
- mHC: Manifold-Constrained Hyper-Connections (DeepSeek), arXiv:2512.24880. https://arxiv.org/pdf/2512.24880 ; https://huggingface.co/papers/2512.24880
- South China Morning Post — "DeepSeek kicks off 2026 with paper signalling push to train bigger models for less." https://www.scmp.com/tech/big-tech/article/3338427/
- Introl — "DeepSeek mHC: The Architecture Fix That Could Unlock [scaling]." https://introl.com/blog/deepseek-mhc-architecture-breakthrough
- CNBC — "Humanoid robotics company raises up to $1.4 billion from Nvidia, Amazon and others," June 10, 2026. https://www.cnbc.com/2026/06/10/neura-robotics-funding-ai-humanoid-robots.html
- Briefs.co — "Robotics Startups Raised $23B in 2026, Close To All Of 2025." https://www.briefs.co/news/robotics-startups-raised-23-billion-in-2026-closing-in-on-all-of-2025/
- Fortune — "Anthropic confidentially files for IPO after raising $65 billion at a $965 billion valuation," June 1, 2026. https://fortune.com/2026/06/01/anthropic-confidentially-files-ipo-965-billion-valuation/
- Sacra — "Anthropic revenue, valuation & funding." https://sacra.com/c/anthropic/
- BuildFastWithAI — "AI News Today - June 15, 2026" (Anthropic revenue trajectory, compute strategy, profitability). https://www.buildfastwithai.com/blogs/ai-news-today-june-15-2026
- TechTimes — "Google Gemini 3.5 Pro Nears June Launch With 2 Million Token Context And Deep Think Reasoning." https://www.techtimes.com/articles/317919/20260606/
- INDmoney — "Oracle Q4 FY2026 Earnings: Why ORCL Stock Fell 10% Despite a Strong Beat." https://www.indmoney.com/blog/us-stocks/oracle-q4-fy2026-earnings-orcl-stock-drop
- Oracle Corp — Form 8-K, FY2026 Q4 results (SEC). https://www.sec.gov/Archives/edgar/data/0001341439/000119312526265848/orcl-ex99_1.htm
- Bloomberg — "Oracle Earnings Pose Next Test for Suddenly Shaky AI Stock Rally," June 10, 2026. https://www.bloomberg.com/news/articles/2026-06-10/oracle-earnings-pose-next-test-for-suddenly-shaky-ai-stock-rally
- BuildFastWithAI — "AI News Today - June 13, 2026" (Gemini 3.5 Pro timing and pricing). https://www.buildfastwithai.com/blogs/ai-news-today-june-13-2026
- Yahoo Finance — "Chip Selloff Hits SOX After Broadcom's [~13%] Drop," June 5, 2026. https://finance.yahoo.com/markets/stocks/articles/chip-selloff-hits-sox-broadcoms-100511793.html
- StartupHub.ai — "Broadcom tumbles 13% as AI chip guidance stalls; Google rallies on $85B expansion bet, SOXX -2.1%," June 4, 2026. https://www.startuphub.ai/ai-news/ai-stocks-daily/2026/ai-stocks-2026-06-04
Disclosures & Disclaimer
This report is general commentary published for information purposes only. It is not investment advice, a recommendation, or a solicitation to buy or sell any security. Parameter is a research publication, not a registered investment adviser or broker-dealer. Views are the publication's own analytical opinions, are subject to change, and may prove wrong. Readers should do their own research and consult a licensed financial professional before acting. The publication and/or its principals may hold positions in securities mentioned. Company facts and figures are drawn from public sources believed reliable but are not guaranteed. © Parameter.
About Parameter
Parameter publishes a daily, independent brief on the most important advancements in artificial intelligence — models, research, compute, and the market that prices them. Provided for information only; not investment advice. © Parameter. All rights reserved.