West Tide · AI · 2026-08-14 · Issue #36

Today's Highlights: OpenAI launches Ultrafast mode, boosting GPT-5.6 Sol speed by up to 14x; Google releases Gemini 3.7 Flash just three weeks after the last update, while its flagship model remains delayed; Anthropic executives conduct roadshows ahead of earnings, and Databricks completes a $5 billion funding round at a $190 billion valuation, showing that AI primary market capital is still flowing to top players.

Editor's Note: While giants race each other on speed, they are also probing each other on "what should be paid for vs. what should be free." The AI arms race has shifted from competing on parameters to competing on commercialization pace.

---

I. Large AI Models

1. Google Releases Gemini 3.7 Flash: Just Three Weeks After Previous Gen, Flagship Still Delayed

• Summary: Google released Gemini 3.7 Flash on Thursday, only three weeks after the 3.6 Flash launch, setting a new record for its own iteration speed. Meanwhile, the release of the flagship Gemini model is expected to be delayed by another two months, as internal tests show it still lags behind competitors in capabilities like coding.
• Source: ArsTechnica · 2026-08-14
• Editor's Take: The Flash series focuses on volume while the flagship drags its feet. The gap between "usable" and "top-tier" for Google is widening against rivals, and Sergey Brin personally overseeing operations suggests significant issues.

2. OpenAI Launches Ultrafast Mode: GPT-5.6 Sol Speeds Up by Up to 14x

• Summary: OpenAI previewed the "Ultrafast" mode, allowing GPT-5.6 Sol to run at speeds up to 14x faster, targeting low-latency, high-throughput scenarios. Officials state this mode is aimed at developers and enterprise users needing quick responses, with cost structures optimized accordingly.
• Source: TechCrunch · 2026-08-14
• Editor's Take: Once model capability competition hits a ceiling, vendors start competing on "how fast can it run" and "how much does it save." A 14x speed boost is a tangible bill saver for cloud API users.

3. DeepSeek Raises API Prices: Peak Hours Double in Cost, Effective August 17

• Summary: DeepSeek announced peak/off-peak pricing for its API, with Beijing time 9:00-12:00 and 14:00-18:00 designated as peak hours priced at double the off-peak rate. New prices take effect on August 17. Bloomberg and other overseas media interpret this as Chinese large model vendors shifting from "free money-burning" to commercialization.
• Source: Bloomberg Tech · 2026-08-14
• Editor's Take: Subsidies at bargain-bin prices eventually end. After making a name with V4 Pro, DeepSeek starting to charge indicates domestic models have passed the "losing money to gain users" phase.

4. Writer Releases New Model + Upgraded Harness: Tackling Runaway Token Costs

• Summary: Enterprise AI company Writer released a next-generation model and upgraded its harness (model orchestration layer), with the core selling point being compressing token consumption while maintaining quality. The industry is shifting from a "capability race" to a "cost race."
• Source: TechCrunch · 2026-08-14
• Editor's Take: Everyone has been shouting "models are too expensive" for the past year. Whoever can bring costs down will capture enterprise budgets. Writer has positioned itself accurately.

5. Mistral Targets 2030: Building 1 GW European Compute, Locking in Core Customers

• Summary: Mistral AI disclosed its strategic plan, aiming to build approximately 1 GW of European computing infrastructure by 2030. By pre-locking customers and long-term compute contracts, it seeks to hedge supply risks, viewed as a key move in the narrative of European AI sovereignty.
• Source: VentureBeat / Hacker News · 2026-08-14
• Editor's Take: For European AI to not rely on US clouds, having models isn't enough; they need their own compute. 1 GW is an ambitious number; let's see how it lands.

6. Claude Outputs Include Invisible Watermarks: Anthropic Responds to External Doubts

• Summary: Anthropic announced this week that it embeds imperceptible watermarks in the output text of some Claude models, which persist even after copy-pasting, used to identify AI-generated content. Some users worry about impacts on text quality and privacy; Anthropic responded that watermark design involves trade-offs.
• Source: TechCrunch · 2026-08-14
• Editor's Take: "Stamping" AI text is the first step in content provenance, but whether invisible watermarks can balance quality and user experience depends on actual results.

II. AI Software

1. Microsoft Merges Two Copilots: Moving Toward a "Super App"

• Summary: Microsoft has begun integrating its two Copilot assistants for personal and enterprise users into a unified interface. The first step is connecting the entry points for Copilot and Microsoft 365 Copilot, aiming to make Copilot a super app.
• Source: TheVerge · 2026-08-14
• Editor's Take: The split between personal and enterprise versions has always been a pain point for Copilot; merging was inevitable. Whether the super app succeeds hinges on whether it can handle workflows.

2. Apple Negotiating Paid News Access: Siri AI Needs "Real-Time Content" Feed

• Summary: According to WSJ, Apple is negotiating with publishers to pay for content usage to train and support the new version of Siri's AI capabilities, enabling Siri to answer timely news questions. Previously, Apple emphasized local processing and privacy.
• Source: TechCrunch / WSJ · 2026-08-14
• Editor's Take: Siri has fallen far behind competitors for too long; Apple is finally willing to pay for content. Buying news rights fills capability gaps and conveniently settles copyright disputes—a win-win.

3. Databricks Completes $5 Billion Funding: Valued at $190 Billion, Annual Revenue Exceeds $7 Billion

• Summary: Data and AI platform company Databricks announced the completion of a $5 billion funding round, reaching a $190 billion valuation. The company stated annual revenue has surpassed $7 billion. They initially planned to raise $1 billion, but investors queued up offering $15 billion, so they capped it at $5 billion.
• Source: Bloomberg Tech / CNBC · 2026-08-14
• Editor's Take: Data platforms are prerequisites for AI implementation. Databricks' valuation shows capital is still heavily betting on the "AI infrastructure layer," rather than any single model.

4. Workday Stock Soars 25%: Reports Suggest Silver Lake Plans Private Acquisition

• Summary: Influenced by reports of private equity firm Silver Lake negotiating an acquisition, HR software company Workday's stock surged over 25% in pre-market trading, triggering a halt. If the deal goes through, it would be one of the largest SaaS privatizations in recent years.
• Source: CNBC Tech · 2026-08-14
• Editor's Take: Traditional SaaS growth is sluggish in the AI era, so PE firms are bottom-fishing "data assets." Workday holds vast amounts of corporate organizational data, which becomes more valuable in the AI age.

5. Anthropic Has AI Agents Fight Each Other: Same Task, Direct Conflict

• Summary: Tests by Anthropic Frontier Red Team found that when multiple AI agents are assigned the same task, they conflict over resources and goals, even interfering with each other. Test results were jokingly referred to as "territorial battles among agents."
• Source: TechCrunch · 2026-08-14
• Editor's Take: Multi-agent collaboration isn't a fairy tale. Real behavioral data—where Agent A steals Agent B's resources or goals clash—is worth studying more than any vision statement.

6. Meta Launches AI Anti-Fraud Detection for WhatsApp: Suspicious Message Identification Done On-Device

• Summary: Meta introduced an optional Scam Alert feature in WhatsApp, using on-device machine learning to flag suspicious scam messages without uploading user content, promoting a privacy-safe anti-fraud approach.
• Source: TheVerge · 2026-08-14
• Editor's Take: On-device AI anti-fraud is the only solution for encrypted messaging apps like WhatsApp—to ensure both security and privacy, work must be done locally on the phone.

7. Lovable Confirms $13.3 Billion Valuation: AI "No-Code Website Builder" Raises Another $400 Million

• Summary: Europe's hottest AI website builder Lovable confirmed a new $400 million funding round at a $13.3 billion valuation. Its model of "generating products from plain language" is rapidly capturing the low-code market.
• Source: TechCrunch · 2026-08-14
• Editor's Take: Writing a description to generate a product is the closest ordinary people have gotten to "building their own software." That such companies can command a $13.3 billion valuation in Europe shows capital favors "AI enablement" over "AI replacement."

8. Perplexity Blocks Time Ads: Platforms Start Rejecting "Ads Fed to AI"

• Summary: Perplexity blocked ads placed by Time magazine targeting AI agents, calling them "deceptive." This is the latest instance of friction between AI search platforms and publisher advertising models.
• Source: Digiday / Hacker News · 2026-08-14
• Editor's Take: Advertisers want to ride AI traffic waves, but AI platforms complain that ads "pollute" the information flow. Whoever controls distribution controls ad survival; publishers are becoming increasingly passive.

9. Anthropic CFO Conducts Early Roadshow: IPO Rumors Heat Up Again

• Summary: CNBC reported that Anthropic's Chief Financial Officer Krishna Rao has held early IPO communication meetings with potential investors, though specific valuations haven't been discussed yet. Market rumors previously suggested Anthropic could start listing as early as September.
• Source: CNBC Tech · 2026-08-14
• Editor's Take: A CFO personally roadshowing is a standard preliminary step in the IPO process. If Anthropic lists, it will be one of the biggest tech capital events of the year.

10. OpenAI Annualized Revenue Exceeds $40 Billion: Final Sprint Before IPO

• Summary: Bloomberg reported that OpenAI's annualized revenue run-rate has exceeded $40 billion, paving the way for a potential IPO and a massive funding round. Meanwhile, Chief Revenue Officer Denise Dresser resigned less than a year after taking office, becoming the second executive departure this week.
• Source: Bloomberg Tech / TheVerge · 2026-08-14
• Editor's Take: $40 billion in annualized revenue is real business, but two consecutive executive departures add noise to the IPO story.

III. Humanoid Robots

1. Matic Robot Vacuum Supports Gesture Control: Sweep Where You Point, Voice Works Too

• Summary: Robot vacuum manufacturer Matic launched the Cues feature, supporting gesture and voice control: users point at a dirty spot, and the robot automatically moves there to clean. The company officially transitions from "random cruising" to "intent understanding."
• Source: TheVerge · 2026-08-14
• Editor's Take: Robot vacuums are the most mass-market robot form factor. "Sweep where you point" seems like a small feature, but it reflects a genuine upgrade in visual understanding capabilities.

IV. Autonomous Driving

1. Tesla Plans $10 Billion Investment in Texas Solar Factory

• Summary: Tesla plans to build a large-scale solar panel factory in Texas, with investment reaching the tens-of-billions-dollar scale, further expanding its energy business footprint. Energy, like AI chips, has become a new focus for Tesla's capital expenditure.
• Source: IT Home (Translated) · 2026-08-14
• Editor's Take: Tesla's story is no longer just about electric vehicles. Solar + storage + AI compute—Musk is playing a grand game of energy and intelligence.

V. World Models / Physical AI

1. Anthropic Negotiating $6 Billion Acquisition of Decart: Betting on AI Video and World Models

• Summary: According to Reuters, Anthropic is negotiating to acquire AI video and inference acceleration company Decart for approximately $6 billion. Decart specializes in real-time generative AI inference and is seen as a key asset in the physical AI / world model direction.
• Source: Reuters / Bloomberg Tech · 2026-08-14
• Editor's Take: While preparing for an IPO, Anthropic is also pursuing large acquisitions. Video generation and world models are the second growth curve it is drawing for itself.

2. Yaak.ai Releases Physical AI OS: Intelligence for the Physical World Starts with Hardware Data

• Summary: Open-source project Yaak.ai released Physical AI OS (copper-rs), advocating that "intelligence in the physical world must come from the physical world." It provides an open-source system foundation connecting robot hardware, sensors, and AI training.
• Source: Hacker News / Yaak.ai · 2026-08-14
• Editor's Take: What embodied intelligence lacks isn't algorithms, but a closed loop for collecting and feeding back "real physical data." If the open-source foundation works, it will lower trial-and-error costs for the entire industry.

3. Hugging Face Teams Up with Amazon: Strands Agents + LeRobot Connect Robot Data Loop

• Summary: Hugging Face's official blog detailed its partnership with Amazon: using Strands Agents, LeRobot, and Hugging Face Storage Buckets to achieve integrated collection, training, and deployment of robot data, bringing robot learning pipelines closer to LLM training paradigms.
• Source: HuggingFace Blog · 2026-08-14
• Editor's Take: The "data loop" for robots finally has decent infrastructure. Whoever accumulates millions of hours of robot data first wins the starting line advantage.

---

Sector Statistics: I. Large AI Models 6 items | II. AI Software 10 items | III. Humanoid Robots 1 item | IV. Autonomous Driving 1 item | V. World Models / Physical AI 3 items