Google DeepMind ships Gemini 3.7 Flash: double-digit gains on coding and agent benchmarks, half the price
Google DeepMind released Gemini 3.7 Flash on August 13, three weeks after Gemini 3.6 Flash, calling it "our most intelligent workhorse model yet for coding and agents." The benchmarks tell a more concrete story than the marketing language: FrontierCode 1.1 Main jumped from 34.4% to 43.6%, DeepSWE v1.1 from 49.0% to 65.3%, WebDev Arena Elo from 1538 to 1588, AutomationBench from 17.0% to 30.4%, GDP.pdf from 22.0% to 34.0%, Terminal-bench 2.1 from 78.0% to 85.8%. The model supports 1M tokens of context and 64K output, with three thinking levels (low/medium/high), and it is the new backbone for Gemini Spark, Google's 24/7 personal agent that runs in 160+ countries.
The pricing is the other half of the announcement, and probably the more disruptive half. The introductory rate is $0.75 per million input tokens and $3.75 per million output — half the original 3.6 Flash price — and that price is locked through December 31, 2026, reverting to $1.50 / $7.50 starting January 1, 2027. For teams running high-volume agent loops where every weak turn triggers a re-roll, halving inference cost while pushing deep agentic numbers up by double digits is exactly the combination that shifts build-versus-buy math, not just a marketing footnote. The model is already live across the Gemini API, AI Studio, Android Studio, Antigravity, the Enterprise Agent Platform, and Spark — though notably not in the standard Gemini chatbot UI for free users.
The cadence is the part I keep coming back to. Three weeks between Flash releases is fast even by 2026 standards, and it puts real pressure on every other agentic model team to either match the velocity or differentiate on something other than the latest benchmark delta. For engineering teams running production pipelines on 3.6 Flash, the upgrade looks easy on paper — same API surface, better numbers, lower cost through year-end — but the practical work is verifying that your existing prompts and tool schemas hold up against a model that is meaningfully better at multi-step planning. The honest caveat, as always: these are vendor-reported numbers, and the gain from 3.6 to 3.7 Flash on long-horizon engineering is large enough that I want to see independent verification before I rebuild prompts around it.
— Google DeepMind (blog) · Google (blog)
🔗 Introducing Gemini 3.7 Flash (deepmind.google) · Gemini 3.7 Flash (blog.google) · Silicon Report coverage · Oton Technology detailed benchmark table
SpaceX closes the $60B Cursor acquisition — the largest startup deal in history
SpaceX officially closed its $60 billion all-stock acquisition of Cursor (legal entity Anysphere) on August 14, confirmed by an SEC Form 8-K filing. A SpaceX subsidiary (X67 Inc.) merged into Anysphere, with Cursor surviving as a wholly owned subsidiary. Cursor's common and preferred shares converted into 389,573,254 SpaceX Class A common shares, plus roughly 29.1 million RSUs and 44.4 million stock options. The implied Cursor equity value of $60 billion was set against SpaceX's 7-day volume-weighted average closing price before closing. This is the largest venture-backed startup acquisition ever recorded, at a 3.4% dilution to SpaceX's post-IPO valuation. If the deal had fallen through, SpaceX had agreed to pay Cursor a $1.5 billion termination fee plus $8.5 billion in computing resources — a $10 billion breakup-fee structure that signals how seriously SpaceX pursued this deal.
Cursor's own announcement kept it focused on one sentence: "We will have access to the largest fleet of GPUs in the world, giving us the compute to build stronger models that are also more economical to run." Cursor hit $2 billion ARR with >1 million paying users, 7 million monthly active users, and deployment at more than half of the Fortune 500. The company's product statement is no longer "AI code completion" — it is now positioned as the AI coding layer inside SpaceX's Grok Build, Grok Bot, and Grok API ecosystem. Grok 4.6, released the day before closing, is the first joint product of the collaboration. The Grok 4.5 launch in July had already moved Cursor from OpenAI/Anthropic model dependence toward xAI's Colossus supercluster for training; the acquisition formalizes that integration.
The competitive dynamics are genuinely unusual, and worth pausing on. Anthropic — one of the rivals this deal is meant to help SpaceX catch — is also a SpaceX compute customer, having agreed in May to pay roughly $45 billion over three years. Google has a similar arrangement. The same infrastructure now serves SpaceX's own models and its competitors' models, which is the business SpaceX is in: selling compute, funding models with it, then absorbing the developer interface layer via Cursor. Musk told SpaceX staff earlier this month that AI revenue would out-earn rockets by September. Whether the AI coding market responds by fleeing to open-source harnesses (which some teams have already done with claimed 97% cost reductions) or by accepting lower Cursor pricing enabled by owned compute is the test that will define whether this $60 billion was a strategic win or a defensive overpay. The Cursor Privacy Mode change — routing session data into Grok model training — is the immediate practical consequence enterprise IT buyers will need to audit.
— SpaceX (SEC Form 8-K) · Cursor (official note) · Bloomberg
🔗 Cursor official note (via bizstack summary) · AlphaSignal deal structure deep-dive · Bloomberg via Europa Press · daily.dev 8-K summary
Databricks raises $5B at $190B — "token maxing has freaked out the CFOs"
Databricks closed a $5 billion strategic funding round at a $190 billion valuation on August 13, led by Coatue along with Blackstone, MGX, T. Rowe Price, and new investor Sixth Street Growth. The timing matters. This comes roughly six months after a $5B round at $134B, which means valuation jumped 42% in half a year while revenue climbed even faster. The company crossed a $7 billion annualized revenue run rate in Q2 with >80% year-over-year growth, >$1.5B run rate for Lakehouse (over 100% YoY), and >$100M run rate for Lakebase, a Postgres-style database launched in 2025.
CEO Ali Ghodsi told TechCrunch that the company had originally set out to raise $1 billion and instead saw $15 billion in demand materialize within days of a news leak about the round, which he summarized as "my phone blew up." The story investors are buying is not just data warehousing. Databricks is positioning Lakebase, Genie (its "AI coworker" agent), and Unity AI Gateway (multi-model governance and cost controls) as the three-product platform that enterprises actually need to deploy agents that don't burn through budgets. Lakebase and Genie run on the same Databricks stack; Unity AI Gateway is where model routing and budget enforcement happen. Ghodsi gave the most quotable reason for the demand surge: rising AI token costs have "freaked out the CFOs," and customers are now actively considering Chinese open-weight models they previously ignored, because the unit economics force the conversation.
The honest read on $190 billion: it prices the company at roughly 27x annualized revenue, aggressive by historical software standards but grounded in specific metrics — positive adjusted free cash flow for 12 consecutive months, 1,000+ customers at $1M+ ARR, 100+ at $10M+, and 70% of the Fortune 500 on the platform. The IPO Ghodsi confirmed is part of the plan, but he downplayed any urgency: "right now I just think there would be too much distraction in the public market." The deeper signal is that the agent infrastructure layer is becoming the place where money goes — and the fact that the company's CEO is openly telling customers to consider Chinese models when cost pressure mounts is a more honest read on the state of the API market than any analyst note.
— Databricks (newsroom) · Bloomberg / Yahoo Finance
🔗 Databricks press release · Yahoo Finance / Bloomberg · CRN on customer and product metrics · TechTimes $15B demand analysis
Lovable raises $400M at $13.3B — vibe coding consolidates
Lovable, the Stockholm-based vibe-coding platform, raised $400 million in Series C funding at a $13.3 billion valuation on August 12, co-led by Menlo Ventures and EQT's Scaleup Europe Fund. Tencent joined as a new investor alongside Balderton Capital, Carmignac, Kaszek Ventures, LTS Growth, World Innovation Lab, and Regent. The round more than doubles the company's December 2025 valuation of $6.6B and roughly 7x's its July 2025 valuation of $1.8B. Crucially, the new investors span three continents, signaling that the platform is being read as a global business opportunity, not just a European AI story.
The commercial metrics are unusually concrete for a vibe-coding startup. Since launching in November 2024, users have built more than 60 million projects on Lovable, and apps built on the platform see over 900 million visits every month. Annual recurring revenue has nearly tripled from $200 million in December and is tracking toward roughly $600 million by the end of August, with the company reporting >1 million paying users and growing enterprise penetration — employees at nearly two-thirds of the Fortune 500 have used the platform, including Adidas, Nvidia, and Deutsche Telekom. The Series B was 8 months ago at $6.6B; ARR was $200M then. ARR is now $400M+ and heading to $600M, which means the valuation step-up from $6.6B to $13.3B is roughly in line with revenue growth, not a hockey-stick multiple expansion.
What I find hard to look away from is the contrast between Lovable's numbers and the broader vibe-coding competitive landscape. Cursor got acquired by SpaceX for $60B a couple of days before this round closed (see separate story), Replit sits much lower, and the no-code platforms of the pre-LLM era (Airtable at $11B in 2021, Bubble, Retool) are now worth less than Lovable's pre-money. Lovable's claim to category leadership is now backed by both revenue and user metrics, but the open question is whether the conversion of free vibe-coders into enterprise contracts holds as the novelty wears off. Anton Osika, the CEO, said the capital will go to product, infrastructure, and security — the boring stuff that determines whether a vibe-coded app survives a Fortune 500 SOC 2 audit. That is the test Lovable has not yet passed at scale, and the next funding round will be priced on whether it has.
— Lovable (official blog) · Reuters via Economic Times
🔗 Lovable Series C announcement · Reuters via Economic Times · BriefAsia on Tencent's B2B shift · MartechAI on revenue and valuation trajectory
NVIDIA puts Spectrum-X CPO into mass production — the network that holds AI factories together
NVIDIA announced on August 14 that its Spectrum-X Ethernet Photonics switch has entered full mass production — the first 200G-per-lane co-packaged optics (CPO) Ethernet switch system to do so, built on the new Spectrum-6 chip. The numbers NVIDIA gives for the platform are striking in the specific way infrastructure buyers care about: laser count reduced to roughly one-quarter of a conventional pluggable-optics design, power consumption to roughly one-fifth, mean time between incidents extended tenfold, optical loss from ~22 dB to ~4 dB, signal integrity improved ~64x. Spectrum-6 itself doubles per-chip bandwidth over Spectrum-4 (102.4 Tb/s vs 51.2 Tb/s). CoreWeave, Lambda, and Oracle are the first hyperscale customers; the supply chain runs through TSMC (silicon photonics), SPIL (advanced packaging), Lumentum and TFC (lasers), and Foxconn (system assembly).
The strategic read is what makes this bigger than a product launch. As AI clusters scale past hundreds of thousands of GPUs, the bottleneck shifts from compute to interconnect: the cost and power of plugging optical transceivers into switch front panels grows linearly with port count, while CPO co-locates the optical engine with the switching ASIC so the electrical path that consumes power is dramatically shorter. NVIDIA's claim that traditional pluggable-optics networks hit ~22 dB of loss and CPO drops that to ~4 dB is the kind of number that determines whether a 100,000-GPU cluster is even buildable at acceptable power budgets. The 2025 launch of Spectrum-X Photonics, the January 2026 inclusion in the Vera Rubin platform, and now mass production form a single arc: NVIDIA is no longer just selling GPUs into AI factories, it is selling the network that holds them together.
The honest caveat is that CPO is not new as a concept, and Broadcom, Marvell, Intel, and several Chinese switch vendors have their own versions. What NVIDIA has that the others do not is a captive demand pool — every AI factory its GPUs anchor is now a CPO customer — and a vertically coordinated supply chain that can manufacture at scale. The procurement question I would ask if I were a hyperscaler: how long do pluggable 1.6T transceivers remain the right choice for new AI cluster builds, and at what GPU-count threshold does the economics flip to CPO? NVIDIA is betting the flip happens sooner than most buyers think.
— NVIDIA AI Infra (official X) · IT之家 via Caixin
🔗 Spectrum-X mass production announcement (Toutiao/IT之家) · Spectrum-X technical breakdown (JUR) · 上游财经/Toutiao on production chain and customer ramp · 观察者网 on losses and bandwidth details
LG and NVIDIA sign humanoid + AI factory + mobility MOU
LG and NVIDIA signed a memorandum of understanding on August 13 at NVIDIA's Santa Clara headquarters, with LG Corp. Chairman and CEO Koo Kwang-mo and NVIDIA CEO Jensen Huang in attendance. The agreement covers three areas, and the most concrete one is robotics: LG will develop a next-generation bipedal humanoid robot using NVIDIA's open Isaac GR00T foundation model, NVIDIA Jetson Thor for onboard compute, and NVIDIA Halos for Robotics (a full-stack safety system). The humanoid targets a public unveiling in Q1 2027. Within this year, LG will deploy its wheeled CLOiD robots at the washing machine production line of LG Electronics' Tennessee factory for real-world POC validation, with expansion to global production sites, homes, and commercial spaces based on results.
The other two pillars are infrastructure-heavy. On AI factories, LG will combine its thermal management, battery, design, and operations capabilities with NVIDIA's DSX architecture to build reference sites using NVIDIA's next-gen Vera Rubin platform, starting with a pilot in H1 2027 and expanding to an 80MW facility in Cheonan, South Chungcheong Province by H1 2028. On mobility, LG will develop an AI-defined vehicle computing platform combining its in-vehicle infotainment and software stack with NVIDIA DRIVE Hyperion. A joint task force of technical and business experts from both companies will run R&D, on-site validation, and commercialization together. LG also plans to use NVIDIA Isaac GR00T while continuing to develop its own in-house Robot Foundation Model.
The strategic read is that LG is positioning itself as the Korean equivalent of Foxconn-plus-Quanta: a vertically integrated manufacturing partner that can ship the full physical stack (actuators via LG Electronics, sensors via LG Innotek, batteries via LG Energy Solution, software and integration via LG CNS) alongside the AI compute that NVIDIA provides. For NVIDIA, the win is that every LG-branded humanoid and AI factory carries its technology stack into the Korean and global markets at a scale a single vendor cannot replicate alone. The honest caveat: this is an MOU, not a product launch, and the only thing in market this year is the CLOiD wheeled pilot at the Tennessee plant. Whether LG's bipedal humanoid can move from prototype to commercial deployment in 2027 depends on how much of the physical AI learning curve it inherits from NVIDIA's simulation and Isaac Sim tooling — and how much it has to discover on its own factory floor.
— LG (via PRNewswire) · Korea JoongAng Daily
🔗 LG official press release on PRNewswire · Korea JoongAng Daily full details · Chosun English · DigitalToday English
Qwen becomes the world's most downloaded open model family
Hugging Face published its "State of Open Models: Summer 2026" report on August 14, and the headline finding is structural rather than incremental: Alibaba's Qwen family of open-weight models recorded roughly 2.045 billion downloads on the Hugging Face Hub in the first seven months of 2026, against 418 million for Google's models and 227 million for Meta's. Across all platforms including ModelScope, Alibaba reports more than 3 billion cumulative global downloads. Qwen-based models now account for 151,448 derivatives on the Hub, 2.6× Meta's total footprint and 4.7× the Llama repositories specifically; Google follows with 82,506 derivatives. New Qwen derivatives are being published at roughly 180-210 repositories per day throughout the first seven months of 2026.
The licensing picture is the more disruptive finding for US labs. Of 178 Chinese open releases above 20B parameters this year, 59% carry Apache 2.0 and 22% carry MIT, with none carrying a non-commercial restriction. DeepSeek and Z.ai ship models between 700 billion and 1.65 trillion parameters under plain MIT. On the American side of the same size band, only 29% is Apache or MIT, 41% sits under custom terms, and 30% declares nothing at all. The size gap is also widening: the monthly upper end for models from Chinese labs ranged from 754 billion to 2.78 trillion parameters in 2026, while the largest US open models remained below 130 billion parameters in five of seven months. Chinese frontier labs — Moonshot, DeepSeek, Z.ai, MiniMax — are also the only accounts on the Hub where the heavy 70B+ band carries the volume: effectively 100% of MiniMax's 2026 downloads, 88% of Moonshot's, 55% of DeepSeek's, 39% of Z.ai's.
What this means for the open-weight market is hard to overstate. Alibaba has released more than 460 Qwen models as open source — from sub-billion variants up to Qwen 3.8-Max (2.4 trillion parameters) — and the breadth of the family is what makes the derivative count compound: a developer can pick the size that fits their hardware and still ship on a familiar base. The honest caveat, which Hugging Face itself flags, is that download counts measure integration into pipelines, not frontier capability. Likes track what is exciting; downloads track what is wired into scheduled jobs. All-MiniLM-L6-v2 (a 2022 model) pulled 1.55 billion times in seven months, while Kimi-K3 pulled only about 60 times per like. The Qwen base-model story is real — but the next chapter will be whether the API and cloud revenue that funds the open releases catches up to the scale of distribution they have already achieved.
— Hugging Face (official report) · Alibaba (official statement via Bloomberg)
🔗 Hugging Face State of Open Models: Summer 2026 · China Daily coverage · Tech in Asia · Bloomberg via Forklog
Top comments (0)