Tech Express | August 4, 2026 DeepSeek V4 Flash Draws the AI "Execution Line"

1. DeepSeek V4 Flash Draws the "Execution Line" for AI Models, Shifting Global Pricing Power Eastward

On July 31, DeepSeek officially released V4 Flash, which only retrained post-training code and agent task capabilities, yet sent massive shockwaves through the global developer community. In a single day, it consumed 8 trillion tokens through the AI coding tool OpenCode alone, exceeding the daily average of OpenRouter's entire platform hosting over 400 models. Behind this number lies the explosive growth of agent tasks—a single task can invoke tools dozens of times, consuming tens or even hundreds of times more tokens than traditional Q&A.

On the pricing front, V4 Flash charges just 2 RMB per million output tokens, while Anthropic's Claude Opus 4.8 costs $25 (approximately 170 RMB)—a difference of about 85x. In Artificial Analysis benchmarks, Opus scored 56 compared to V4 Flash's 50, but the API fees for running the same evaluation suite were 52 times higher.

The industry now calls this the "DeepSeek Execution Line"—anything more expensive isn't significantly better, and anything weaker isn't cheaper. This line is systematically pushing mid-tier models lacking irreplaceable value out of the default invocation position, marking an irreversible eastward shift in global AI model pricing power.

2. Amazon Accidentally Burned $1.8 Million Using AI for Coding, 860% Over Budget

According to the Financial Times, multiple internal AI deployments at Amazon have resulted in "catastrophic" cost overruns. In the most severe incident, a team used Claude Sonnet to match author details with product listings, producing massive matching errors that resulted in a final bill of $1.8 million—860% over the allocated budget.

Previously, such mistakes were "trivially cheap," but AI models have made every error "catastrophically expensive." The root cause lies in the lack of effective cost control mechanisms and AI usage monitoring—engineers could casually invoke expensive large models without any spending limits. Amazon is now urgently drafting internal AI usage guidelines to rein in spiraling AI expenditures. This incident serves as a wake-up call for the entire industry: AI may be powerful, but uncontrolled invocation costs can devour any budget.

3. AI Microdramas Fuel "Face-Selling" Economy: Up to $700 Per Face

The explosive growth of China's AI microdrama market has spawned an unusual industry chain—ordinary people licensing their facial likenesses to AI production companies for generating digital characters. On platforms like ActID and New Claw, licensing fees range from $15 to $700, with platforms taking a 10% commission. Since early 2026, over 10,000 AI-generated animated microdramas have been released each month, with the comic-style market reaching approximately 16.8 billion RMB ($2.3 billion) in 2025.

However, privacy and portrait rights disputes have followed. The Guangzhou Internet Court has handled about 700 AI-generated likeness disputes over the past three years. The Actor Committee of the China Radio and Television Association warned that AI face-swapping and voice cloning infringe performers' rights. Once a face is uploaded into an AI system, can its owner ever truly regain control over their digital identity?

4. SemiAnalysis Report: NVIDIA Rubin Ultra Storage Downgraded, HBM Reduced to 192GB

Research firm SemiAnalysis released a report indicating that NVIDIA's Rubin Ultra, previewed to key clients, features specifications lower than market expectations. The core SKU still uses HBM4, but the HBM configuration has been reduced from the previous 12-Hi 288GB to 8-Hi 192GB, with chip power consumption around 1800W. The report attributes the downgrade primarily to tight HBM supply and data center power constraints.

However, Rubin Ultra's main upgrade focus has shifted to NVL576-level large-scale interconnection—8 racks of 72 GPUs each connected through NPO for cross-rack communication. SemiAnalysis believes this helps reduce material costs, eases next year's HBM price increase pressure, and improves deployment efficiency in terms of FLOPs per watt, better aligning with current data center power environments.

5. DDR4 Hits 10-Year Price High, Average Reaches $24 Per Chip

According to DRAMeXchange data, the average spot price for mainstream PC DDR4 8Gb memory chips reached $24 (approximately 163 RMB) in July, up 14.3% from June, marking a historical peak since the agency began tracking this category in 2016. At the start of the year, the average price was just $11.50 (approximately 78 RMB), representing a cumulative increase of 109% in half a year. Compared to the $2.90 baseline from the first 2016 measurement, prices have surged over 8x.

NAND flash prices are rising just as dramatically, with 128Gb MLC NAND chips seeing a 218% increase in half a year. TrendForce expects Q3 traditional DRAM contract prices to rise 13%-18% sequentially. With AI server and HBM demand continuing to expand and squeezing traditional DRAM capacity, Samsung, SK Hynix, and Micron maintain significant bargaining advantages in this supply-constrained environment.

6. Apple Plans to Open iPhone Clipboard Access to Windows PCs, Earliest Fall 2027

Apple has confirmed on its EU Interoperability Requests page that it will develop iPhone clipboard sharing for Windows PCs. This decision stems from Microsoft's request filed in March, asking for persistent clipboard synchronization similar to Universal Clipboard, enabling seamless copy and paste between iPhone and Windows PC.

According to Apple's June 26 update, the feature will be implemented through a framework similar to Accessory Notifications, requiring users to grant consent once for each paired Windows PC. Apple described this as a "significant engineering effort," with completion planned for fall 2027 and a developer beta release shortly after. This suggests the feature could debut as part of iOS 28, finally breaking the long-criticized "ecosystem wall" between iPhone and Windows.

7. Google Pixel 11 Pre-Orders Move Up to August 12, Pro Fold Officially Teased

Google has released back-to-back teaser videos for the Pixel 11 Pro and Pixel 11 Pro Fold, while the Google Store homepage now displays a countdown pointing to August 12 at 7 a.m. PT. This means pre-orders will open hours before the Made by Google 2026 event at 3 p.m. PT that same day. Best Buy has also launched its Pixel 11 series pre-order landing page.

Leaks reveal the Pixel 11 Pro Fold features a new pill-shaped camera bar design, with official retail availability starting August 20. This marks the first time Google has moved pre-orders significantly ahead of the launch event, highlighting its growing confidence in the Pixel 11 series. Full specifications and pricing have also been exclusively leaked by Android Headlines.

8. Rubin Observatory's 3.2-Gigapixel Camera Captures 650,000 Galaxies in a Single Frame

The Vera C. Rubin Observatory has used the world's largest digital camera—the 3.2-gigapixel LSST Camera—to photograph the famous COSMOS field, revealing over 650,000 galaxies and more than 50,000 stars in a single image. The COSMOS field, located in the constellation Sextans covering an area roughly three times the size of the full Moon, faces away from the crowded plane of the Milky Way, making it a prime window for observing the distant universe.

This new image not only complements previous observations by the Hubble and Webb space telescopes but also marks the beginning of Rubin's decade-long Legacy Survey of Space and Time (LSST). The observatory will capture a wide-field snapshot of the southern sky every few nights, helping scientists study dark matter, dark energy, supernovae, and near-Earth asteroids. The final dataset will contain billions of objects with trillions of measurements.

9. Samsung Ships 60.5 Million Units in Q2, Holds Global Top Spot as Chinese Brands Decline

According to the latest market data, Samsung maintained its global leadership in Q2 2026 with 60.5 million shipments and a 22% market share. Apple followed closely with 20% share and 24% year-over-year shipment growth. However, Chinese smartphone brands collectively experienced double-digit shipment declines, with Xiaomi, OPPO, and Vivo all posting significant drops.

Meanwhile, Samsung is reportedly considering adopting Chinese-made DRAM for entry-level and mid-range models to reduce costs. ChangXin Memory Technologies (CXMT) debuted on Shanghai's STAR Market on July 27 after raising 57.9 billion RMB ($8.6 billion), becoming Asia's largest IPO of 2026. China's semiconductor self-sufficiency push is accelerating, posing profound challenges to the traditional memory market landscape.

10. ByteDance's Doubao Keyboard Enters AI Voice Race, LLMs Reshape the Input Method Battlefield

ByteDance's newly launched Doubao Keyboard leverages the same Seed-ASR speech model powering its Doubao AI assistant, delivering fast, low-latency speech-to-text. Meanwhile, Sogou Input Method has announced a full AI upgrade, iFlytek has embedded its latest Spark voice model into the text field, and Baidu Input Method has transformed into an "AI creation assistant" powered by ERNIE Bot—major tech companies are compressing their latest LLM capabilities into this "small tool" that gets tapped dozens of times daily.

MobTech Research notes that large models are driving input methods from "passive recording tools" to "proactive intelligent systems." For ByteDance, Doubao Keyboard is the key entry point connecting content, social, and AI services; for Tencent, WeChat Input Method shoulders the responsibility of guarding the social empire's competitive moat. The ancient input method arena is becoming a new battleground for AI ecosystem control.