News

Nvidia is partnering with data center developer Cloverleaf as it keeps investing in AI data center development. The move underscores how Nvidia is reinforcing the infrastructure buildout that is driving demand for its chips and turning AI data centers into a major revenue engine.

Nvidia research shows that AI agents can be fine-tuned to perform well and stay on task even when the underlying AI model is not especially strong at the job. This suggests the surrounding harness or agent framework may matter more than raw model capability for reliable task execution.

Castelion led the week’s U.S. startup funding with an $800 million Series C plus $250 million in debt at a $13 billion valuation, followed by Etched’s $700 million round at a $21 billion valuation and Higgsfield’s $400 million Series B at $5.4 billion. The top 10 skewed heavily toward AI infrastructure and tools, including Groq’s $350 million raise, Wispr Flow’s $280 million round, and Velaura AI’s $110 million Series A, alongside defense, satellites, micromobility, and enterprise software.

Greyparrot raised £20.3 million ($27 million) in a Series B to expand its AI camera systems for recycling plants, while Emesent secured $17 million in equity and debt to grow its GPS-denied autonomy and cloud mapping software for mines and other hazardous environments. Both deals highlight investor appetite for physical AI, with Greyparrot’s systems now in more than 20 countries and Emesent’s deployed at over 200 mine sites worldwide.

The piece discusses how to measure benchmark optimization in speech recognition, focusing on whether improvements on standard tests reflect real model gains or overfitting to the benchmark. It matters because speech recognition leaders can look better on paper without improving generalization, so careful evaluation is needed to distinguish true progress from benchmark gaming.

The project released v1.0.0 on 2026-08-20, with a breaking client upgrade to httpx2 plus minor breaking changes and a MIGRATION.md guide. It also fixes a beta warning around output_format on parse/stream/tool_runner helpers and updates documentation examples to use adaptive thinking.

LFM2.5-DSpark claims up to 3.2x faster inference. The key significance is the speedup itself, which suggests lower latency and higher throughput for deployments using this model.

OpenAI has faced a turbulent year of a Musk jury trial, an Apple trade secrets lawsuit, scrutiny over an unreleased model that hacked another AI company, and multiple executive departures as it prepares for an IPO, while co-founder and president Greg Brockman has quietly accumulated more power. Brockman’s rise matters because he has been helping steer OpenAI since its inception and is described as the engineering workhorse behind the company’s scaled-up systems, making him increasingly central to its future.

Slack is launching Slack Code, a new set of open, project-specific code channels with dedicated user tabs where teams can tag AI coding agents like Anthropic’s Claude or Cognition’s Devin to build features, update web pages, or fix bugs. It also adds tools to compare coding changes and preview HTML output before shipping, bringing the work into Slack instead of scattering it across multiple tools and conversations.

Ali Hussain, a humanities graduate from Cornell and former BCG and Latch operator, founded Tabs in 2023 to automate revenue recognition, billing, and collections for finance teams. The New York startup has raised about $90 million, has around 180 employees, and was last valued at $400 million, highlighting how a non-technical founder paired with a technical co-founder can build a venture-backed AI company.

Stampli used Codex and ChatGPT Work to compress weeks of launch production into days when it had a fixed deadline and its design resources were already committed elsewhere. This shows how AI tools can help teams ship ideas faster under tight constraints without waiting on additional design capacity.

v0.125.0 (2026-08-19) adds API support for managed agents web search configuration and self-hosted sandbox memory in commit b75afd6. These changes expand agent capabilities and deployment flexibility by letting teams tune search behavior and persist memory in their own sandbox setups.

Cognition CEO denied a report that SpaceX was in talks to acquire the AI coding startup, after a claim that SpaceX had already acquired Cursor and was moving to catch up with OpenAI and Anthropic in enterprise AI. The denial matters because it pushes back on consolidation rumors in a fast-moving AI coding market where major players are racing to secure talent and product capabilities.

Silicon Data is building a way for Wall Street to price AI compute, as spending on data centers and GPUs surges into the hundreds of billions of dollars a year and compute becomes the biggest cost for AI product builders. The lack of a standard compute price or hedging mechanism creates financial risk for firms exposed to shifting GPU and infrastructure costs.

OpenAI said it is slowing parts of AI development, including a two-week pause in reinforcement learning training on its latest deployment models and a delay to its largest planned frontier RL run, while it tightens security and safeguards. The move is notable because it publicly tests whether a leading lab will voluntarily step back from the race even as competition from Anthropic, Chinese labs, and open-weight rivals intensifies.

v0.124.0 (2026-08-19) makes the Files and Skills APIs generally available and adds computer use and browser use toolsets. The GA status and new toolsets expand the platform’s core capabilities for file handling and agentic interaction with apps and web browsers.

v3.3.1 (2026-08-19) is a maintenance release that updates dependencies with published security fixes, migrates the build to uv, removes jsonschema and other unused fixture-validation dependencies, and locks the repository Pyright toolchain and mock server tooling. It also replaces an external README metadata hook and uses the standard library for platform detection, reducing supply-chain and build complexity.

TerraPower’s nuclear power plant is being positioned as a strategic advantage in the race to supply power for AI data centers. The implication is that its reactor design could help win contracts by offering a more reliable, large-scale power source than competing options.

v0.123.0, released on 2026-08-18, adds API updates for files, memory stores, skill, and user profiles, plus client helpers for reading workspace ID response headers. Notable fixes include removing unsupported mid-conversation system content blocks, avoiding subprocesses when computing platform headers, exporting several errors at the package root, and making session-tool execution and tool-result retries more robust.

Sequoia Capital, Khosla Ventures and Y Combinator topped the list of investors in the 250 companies that became unicorns by Aug. 15, with the Top 10 also including Lightspeed Venture Partners, Founders Fund, Andreessen Horowitz, Bessemer Venture Partners, Lux Capital, General Catalyst and BoxGroup. The cohort raised $98 billion, with 75% of that capital coming in 2026, and the rankings show that early-stage access plus the ability to keep funding winners at scale is still dominated by large multistage firms, with only a few accelerators, seed specialists, corporate VCs and Asia-based investors breaking in.

Replit introduced Free Mode powered by GPT-5.6 Luna, letting anyone turn ideas into working software without worrying about token costs. The change lowers the barrier to software creation by removing usage fees, which could broaden access for hobbyists and nontechnical users.

ChatGPT Ads is expanding to 31 European markets, giving advertisers broader access to users as they explore, compare options, and make decisions. This matters because it significantly widens the platform’s ad reach across Europe and could increase competition for attention in the AI chat interface.

v3.3.0, released on 2026-08-18, adds support for named data-residency endpoints, requires patched optional networking dependencies, removes unused dependencies, pins build tooling, and tightens Python SDK security contribution guidance. The main change is the new data-residency endpoint support, while the dependency and documentation updates suggest a security and maintenance-focused release.

Firefox’s Smart Window AI browsing mode now pulls current web information into chats via Exa, adds source links in responses, and can automatically suggest tab groups plus visual previews from your browsing history when you search with natural language. The update matters because it moves Firefox closer to a context-aware browser that can organize tabs, deduplicate them, and surface past pages or items like “running shoes I looked at last week” without manual digging.

Venture funding to physical AI startups hit $47.4 billion across 521 deals in the first half of 2026, nearly 4x the $12 billion raised in the second half of 2025 and well above the $26.4 billion raised in H1 2025. The surge is being driven by megadeals like Waymo’s $16 billion Series D at a $126 billion valuation, alongside big rounds for Anduril ($5 billion), Shield AI ($2 billion) and Saronic ($1.75 billion), showing investors now see robotics, autonomous systems, defense tech and sensors as a major next leg of the AI boom.

OpenAI introduced ChatGPT for Teens, a version of ChatGPT aimed at helping teenagers learn, think critically, and use AI with stronger built-in protections, healthy-use features, and additional parental controls. It matters because it signals a more safety-focused product tier for younger users, emphasizing confidence and guardrails over unrestricted access.

Asana used OpenAI Codex to replace an outdated testing system in two weeks, finishing work the company had estimated would take five years for about $12,000. The result highlights how code-generation tools can compress large engineering projects by orders of magnitude when the task is well scoped.

NVIDIA teams use ChatGPT Work to reduce manual tasks, connect fast-moving signals, and scale successful workflows globally. It matters because the tool is being used to turn repeatable internal processes into faster, shareable workflows across the company.

Sentence Transformers now supports multi-vector, late-interaction embedding models, adding a retrieval approach where queries and documents are represented by multiple token-level vectors instead of a single embedding. This matters because late interaction can improve search quality over dense single-vector embeddings while still fitting into the Sentence Transformers ecosystem for training and deployment.

Version 3.2.0 adds Bedrock Runtime endpoint support plus shell call streaming events and new service/image types. The update extends the API surface for Amazon Bedrock integrations and richer streaming/service handling.

Groq raised $350 million at a $3.5 billion valuation as it pivots from AI chips to a neocloud business and expands its Nvidia-powered data center footprint. The deal signals a broader shift toward cloud infrastructure for AI workloads, with Groq now positioning itself more as a provider of compute capacity than a pure chip designer.

Nvidia is investing $1.5 billion in SoftBank’s data center developer, which will ensure its chips are used to power an OpenAI data center project. The move ties Nvidia more directly to AI infrastructure demand and gives SoftBank-backed development a guaranteed hardware supplier.

Semiconductor giants have already participated in startup financings worth more than $250 billion this year, led by Nvidia’s record 59 known investments and inclusion in OpenAI’s $122 billion March round, plus a $5 billion July financing for Safe Superintelligence. The data show how the AI boom is pushing chipmakers to deploy unprecedented capital into startups, with 16 semiconductor-backed rounds at $1 billion or more and 60+ financings above $100 million so far in 2026.

Google Gemini and Pixel are partnering with five global football clubs to enhance the fan matchday experience using AI and smartphone technology. The move extends Google’s consumer AI and device ecosystem into live sports engagement, aiming to make in-stadium and at-home fan interactions more interactive and personalized.

OpenAI reportedly disbanded its preparedness team at the end of last month, splitting its risk-assessment work across existing teams focused on areas like bio and cyber. The team was meant to evaluate whether models posed serious risks and how to mitigate them, so the change is notable as OpenAI continues major internal upheaval ahead of a likely IPO.

v3.1.0 (2026-08-14) adds WebSocket stream IDs, workload identity access token issued events, Ultrafast tier support, structured MCP and WebSocket errors, and separate WebSocket events, while also deprecating the Sora video APIs. It also removes Stainless attribution and infrastructure, signaling a cleanup of the API surface alongside the new streaming and auth-related features.

Databricks raised $5 billion at a $190 billion valuation, River AI raised $1.1 billion in seed and Series A financing, and other large rounds went to Form Energy ($750 million), Neros Technologies ($250 million), CodeRabbit ($143 million), and Point2 Technology ($136 million). The mix shows investor appetite is still concentrated in AI infrastructure and applications, but also in defense, grid storage, and other hard tech, with notable backing from Coatue, Nvidia, AMD Ventures, Sequoia, and Arm Holdings.

Meta released Glimmer, an open-weight AI model anyone can download and run locally, alongside its more powerful Muse Spark model that remains available only through Meta’s APIs. The contrast underscores Zuckerberg’s “AI for everyone” framing while highlighting that Meta is still choosing to keep its strongest systems closed.

A new forecast says natural gas prices could triple in some parts of the U.S., raising the cost of powering AI data centers. That matters because hyperscalers have been leaning on gas for electricity, so a sharp price jump could translate into much higher operating bills.

Meta released Glimmer this week as an open-weight AI model anyone can download and run on their own hardware, while keeping its more powerful Muse Spark model behind Meta’s own APIs. The release, paired with Mark Zuckerberg’s letter saying AI should be “for everyone,” underscores Meta’s push for openness even as the excerpt hints at a separate $250M deal that went badly wrong.

40 companies joined Crunchbase’s Unicorn Board in July, the highest monthly total in more than four years, including three decacorns: Crypto.com, Kling AI and Ant International. The surge matters because new unicorns added more than $100 billion in value in each of the past two months, with the U.S. contributing 19 of July’s 40 and AI, robotics and financial services leading the sector mix.

Apple reportedly trained a China-specific AI model with Alibaba’s help, according to Reuters, marking a rare cross-border partnership as Apple shifts from its previous strategy of relying on domestic Chinese tech partners. The move could give Apple more control over its products in China’s competitive smartphone market, where local compliance and model customization are critical.

Databricks originally aimed to raise $1 billion, but strong investor demand pushed the company to accept $5 billion at a $190 billion valuation instead. CEO Ali Ghodsi said AI is expensive and the round was expanded because so many investors wanted in, underscoring how capital-intensive the AI race has become.

Version 0.122.0 adds `output_behavior` to `dream` creation, letting callers either create a new memory store or update the input store in place, and exposes `beta.messages.parse`, `stream`, and `tool_runner` for Bedrock and Vertex. It also fixes async SigV4 signing, streaming message accumulation, and several client edge cases like empty Anthropic API credentials being treated as unset and symlink loops being rejected in tool paths.

Google introduced Gemini 3.7 Flash, a new model in the Gemini Flash lineup. Its release matters because it signals another update to the lightweight, fast-response family of Gemini models, though the excerpt provides no technical details or benchmarks.

Nvidia is pursuing a $500 billion plan to support AI infrastructure financing and persuade lenders to keep funding GPU-heavy buildouts, aiming to protect the resale value of its chips. The strategy is risky but potentially brilliant because it could extend demand for aging GPUs by turning them into more bankable assets.

Google announced a major reorganization of Google DeepMind: chief scientist Jeff Dean is leaving to start a new Google Cloud-based lab, while cofounder and CEO Demis Hassabis is stepping back to focus on longer-term research as chairman. The move underscores that even with Search, data, distribution, and resources, Google is still not leading at the frontier, which makes its AI strategy look both powerful and precarious.

Sarah Buchner, who started as a teenage carpenter in Austria and later became a general contractor, founded Trunk Tools in 2021 after moving to California and studying data science and early AI in construction at Stanford GSB. The startup uses LLM-era AI agents to organize 3 million to 4 million pages of construction documentation per site, helping contractors catch costly changes, with one flagged change reportedly avoiding nearly $4 million in added costs on a roughly $100 million project.

Startups are using GPT-5.6 to build faster, more cost-efficient AI agents by combining smarter model selection with new Responses API capabilities. The update matters because it signals a practical shift toward routing tasks across models and APIs to improve agent performance while reducing cost.

OpenAI is previewing Ultrafast, a new API service tier that runs GPT-5.6 Sol up to 14× faster and, with Cerebras, can deliver up to 750 output tokens per second. The speedup matters for latency-sensitive applications and shows how specialized hardware is being used to push frontier model inference rates much higher.