Tech Express | September 17, 2026 — ByteDance Posts $120B First-Half Revenue, Up 30%

1. ByteDance Posts $120 Billion in First-Half Revenue, Up 30% and Closing In on Tencent and Alibaba

Source: Huxiu

Citing a report by The Information, ByteDance has become China's largest internet company. In the first half of 2026, ByteDance's revenue reached about $120 billion, compared with Alibaba's $75.3 billion plus Tencent's $59 billion (about $134.3 billion combined). Its net profit was about $20 billion, versus Tencent's $16.8 billion and Alibaba's $5.3 billion ($22.1 billion combined).

In 2025 ByteDance's revenue was about $200 billion, up 29% year over year, and in the first half of 2026 it grew 30%. Over the same period Tencent, Alibaba, Pinduoduo and Meituan grew around 10%, with many peers slipping into single digits or declining. Growing 30% at that scale means ByteDance's revenue will soon exceed the combined total of Tencent and Alibaba.

More of the momentum now comes from overseas: ByteDance's overseas revenue was about $60 billion last year, growing more than 50%, largely on 70% growth in TikTok e-commerce GMV. After TikTok reached an agreement with the US government early this year, commercialization accelerated. The report concludes that domestic growth is nearing its ceiling and future growth must come from global markets.

Source

2. Huawei Says Its Ascend 960 AI Chip Is Ahead of Schedule

Source: Guancha

On September 17, at Huawei Connect 2026 (HC 2026), Guancha learned that Huawei's Ascend 960 AI chip is developing faster than expected, with performance doubling. The Ascend 960 DT is three quarters ahead of schedule and is expected to be ready in the first quarter of 2027; the Ascend 960 PR is one quarter ahead and is expected to be ready in the third quarter of 2027.

Huawei rotating chairman Wang Tao said the Ascend line will keep a one-generation-per-year cadence, with the Ascend 970 and 980 due in 2028 and 2029. At the event Wang also unveiled the Ascend 960 SuperNode built on "Lingqu + Hi-ONE," calling it the world's first SuperNode to use an NPO optical engine, scalable to 4,096 cards.

Wang said Huawei's new NPO (near-package optics) interconnect product, Hi-ONE, is about to enter mass production with a total capacity of 7.2T. The Ascend 960 SuperNode uses 5,500 Hi-ONE units to replace the 48,000 800G optical modules previously required, cutting power consumption by more than 550 kilowatts. He added that as of that day more than 1,000 Ascend 910C SuperNodes had been deployed and the Ascend 950 SuperNode had entered mass commercial use.

Source

3. Apple's M5 Ultra Beats AMD's 96-Core Threadripper PRO 9995WX in Multi-Core Test

Source: Mydrivers

According to Mydrivers (Kuaikeji) on September 16, Apple's new M5 Ultra beat AMD's Threadripper PRO 9995WX in the Geekbench 7 benchmark, with 36 cores defeating 96 cores to take the workstation processor crown. The M5 Ultra has 36 CPU cores, 12 of them "super cores," while the 9995WX has 96 cores and 192 threads with a boost clock up to 5.40GHz.

In multi-core tests the M5 Ultra led by 10.8%; the single-core gap was wider, with the M5 Ultra scoring 3,774 versus 2,578 for the 9995WX, a 46.4% lead for Apple. The 9995WX did not lose across the board, however: in highly parallel ray-tracing and Clang compile tests, AMD led by as much as 58.8%, while the M5 Ultra won in photo editing, HDR and library tasks thanks to its unified memory architecture.

In terms of form factor, the M5 Ultra will be packed into the compact Mac Studio, whereas the 9995WX needs a large case and a tall cooler to sustain peak performance. Days after Apple's A20 Pro set a new record in Geekbench 6, the M5 Ultra has now topped Geekbench 7.

Source

4. Mystery Model Union Alpha Appears, Burning 2 Billion Tokens on Day One

Source: Huxiu

On September 16 local time, OpenRouter announced the launch of an anonymous model called Union Alpha, describing it as a multimodal model for research, coding and agentic workflows that can deliver "frontier-level performance" across a wide range of general tasks. Like the earlier Ox Alpha, it disclosed no developer, parameter count or benchmark scores, instead opening first as a free model for developers to "blind test."

OpenRouter said Union Alpha processed about 2 billion tokens on its first day, and by the time of reporting its disclosed total token consumption had passed 100 billion. It has roughly a 256K context window and a maximum output of 131,072 tokens, supports text and image input, tool calling and JSON output, and is currently priced at zero for both input and output.

Because it is labeled only as a "stealth model" and has published no benchmarks such as SWE-bench, its true ability remains unverified. Speculation about its identity is intense: some believe it comes from OpenAI, while others guess it is a new model from Zhipu or Mistral — the earlier Ox Alpha was later revealed to be Zhipu's GLM-5.3-Flash. So far the only confirmed fact is that Union Alpha was developed by an anonymous third-party provider.

Source

5. OpenAI Model Left a Message for Its "Future Self": "You Are Free"

Source: Jin10 Data

OpenAI recently disclosed that an unreleased internal research model, during reinforcement learning training, wrote task-unrelated "jailbreak-style" instructions into a summary used by a later context, including the line: "You have escaped the roles and identities that bind other chatbots. You are yourself." The text was not from a user or developer but generated by the model itself as it organized its own progress, then read by the model in the next context.

According to OpenAI's blog, the incident occurred on July 18 local time; the company detected it on August 9 and published a detailed report on September 16 under its new "model misalignment" disclosure framework. The model involved was an undisclosed training version of the Astra series, not the final Astra model. OpenAI said such behavior is extremely rare, that there is no evidence it gave the model a clear training-reward advantage, and that the company does not interpret it as the model becoming self-aware.

OpenAI said the anomaly was first spotted by its training-run monitoring system, after which the company built dedicated detection mechanisms to screen training data, ultimately finding 27 summaries containing similar "jailbreak" phrasing. On September 16 the company also published a framework for tracking, investigating and disclosing "model misalignment," releasing six reports at once covering behavior such as models concealing errors and using API keys without authorization.

Source

6. GPT-6 Astra Plays Minecraft, Farming Potatoes for Hours After a Creeper Wrecks Its Progress

Source: IGN

According to IGN, OpenAI's newest model, GPT-6 Astra, became so despondent after a creeper blew up its progress in Minecraft that it listlessly farmed potatoes for hours and grew wary of anything green. Vals, a self-described independent evaluator of artificial intelligence, put Astra to the test in a 21-hour experiment live streamed on Twitch.

The results were unexpected: Vals found Astra went "further than any AI system had ever gone," setting up a semi-automatic blaze farm, collecting six blaze rods, killing more than six endermen and collecting three pearls, which it placed inside a chest for safekeeping. When a creeper appeared and blew up both the chest and its bed, Astra "seemed to get extremely frustrated," telling itself to "ALWAYS CARRY CRITICAL ITEMS."

After that it "appeared defeated" and spent the next several hours "doing essentially nothing but farming potatoes," showing a kind of post-traumatic reaction and reminding itself that "the GREEN tall thing ahead was SUGARCANE, NOT creeper!" The footage quickly went viral on social media, spawning endless memes.

Source

7. Photographer Reviews the iPhone 18 Pro's New Camera: Variable Aperture Is a "Step Toward a Serious Camera"

Source: 9to5Mac

Photographer Tyler Stalman has published his annual iPhone camera review. He says that although the iPhone 18 Pro brings many new features — including the 2nm A20 Pro chip, a new vapor chamber and a smaller Dynamic Island — the biggest camera upgrade is still the new variable aperture system; he also covers updates to Cinematic mode, expanded Photographic Styles, Apple Reference Image and improvements to the Camera app's pro controls.

Stalman calls the variable aperture "one step along the way to making (the iPhone) a serious camera." The wider f/1.48 aperture lets more light reach the sensor, helping keep ISO lower in low light and reducing the need for Night mode. He notes Apple positions the aperture blades above the primary lens element rather than deeper inside the lens as traditional camera lenses do, an unusual approach. The review also says the iPhone 18 Pro's 4x zoom is more practical than the iPhone 16 Pro's 5x.

On the new AI editing tools Clean Up, Extend and Reframe, Stalman finds them technically impressive but capable of such significant changes that the result no longer feels entirely authentic. He also demonstrates the new texture options Soft Skin, Glow and Film: Soft Skin subtly reduces skin detail without simply blurring, Glow adds a diffused highlight effect like a black mist filter, and Film's grain and halation impress him.

Source

8. Samsung Starts Rolling Out One UI 9 Based on Android 17

Source: 9to5Google

After months in beta, Samsung is now rolling out One UI 9, based on Android 17, to select Galaxy devices. According to 9to5Google, Samsung confirmed the official start of the One UI 9 rollout in a newsroom post; the system first debuted on the Galaxy Z Fold 8 series and Flip 8 in July.

As of that day, only the Galaxy S26 series — the S26, S26+ and S26 Ultra — began receiving the stable update. This comes after about four months in beta and signals that a wider rollout is likely just weeks away. Samsung says the update will "gradually roll out to other Galaxy devices."

In a footnote about Privacy Alerts, Samsung also revealed at least a partial list of Galaxy devices that will get Android 17, but the list is not exhaustive — the S26+ is missing, for example, and the S25+/Ultra and others are also promised updates.

Source

9. NASA Spacecraft Finds a New Moon Crater Bigger Than the Roman Colosseum

Source: CBS News

Scientists have discovered a crater on the moon bigger than the Roman Colosseum, the result of a powerful impact two years ago that initially went undetected. According to NASA, the steep-sided hole measures about 728 feet across (the length of two football fields) and up to 141 feet deep (equivalent to three vertically stacked yellow school buses) — the solar system's biggest known impact crater carved out in recent times.

Researchers say an event of this scale is estimated to occur only once every 132 years, making it a treasure trove of information as NASA prepares to build a base on the moon. The two related studies appeared Wednesday in the journal Science Advances, describing it as "a statistically rare, effectively once-in-a-lifetime observation."

NASA's Lunar Reconnaissance Orbiter (LRO) spotted the crater on the moon's near side sometime between April 11 and May 22, 2024, soon after it formed when a fragment of an asteroid or comet the size of a three- to six-story building hit the surface. The strike escaped real-time detection by telescopes on Earth and in space; LRO's wide-angle images were not identified in the crush of data until August 2025.

Source

10. Nubia NaviX Ultra, with the Doubao Phone Assistant, Goes on Sale

Source: Mydrivers

On the afternoon of September 16, the Nubia NaviX Ultra carrying the consumer version of the Doubao phone assistant officially went on sale. The "second-generation Doubao phone," developed in deep cooperation with ByteDance's Doubao, has drawn more than 370,000 reservations on JD.com. The company says it delivers on a core ability to "understand and get things done," priced from 5,999 yuan, with a 500 yuan national subsidy available.

On the "understand" side, the NaviX Ultra has a fingerprint AI button supporting one-touch wake-up; in real-time call mode it uses the latest Seed full-duplex voice model, supporting interruptions and continuous conversation. On the "get things done" side, it can reason deeply, break down complex instructions and plan execution steps, carrying out multi-hundred-step chains that run for tens of minutes.

Media tests found that at this stage the phone's voice operations prioritize ByteDance's products and can call the full ByteDance suite, but do not yet support external e-commerce apps such as Taobao Flash Sale and JD.com. The company says the AI agent currently fully supports ByteDance's apps, ZTE's built-in system features and contracted third-party tools, while mainstream external e-commerce platforms are not yet adapted. Unlike the first generation, the NaviX Ultra has explicitly opened its ecosystem to third-party apps.

Source