Government and private sector contribute roughly equal amounts to India's research and development

Intel and Lens Technology have signed an agreement on a strategic partnership aimed at accelerating the development of solutions for glass substrates. This step was taken on July 25th, three years after Intel first identified the glass substrate as a promising path for advanced packaging.

As part of this collaboration, companies Triassic (Guangdong) Technology and Tongge Micro independently overcame the fundamental problem of brittleness by employing material selection methods, process optimization, structural reinforcement, and control over the entire production cycle.

Yui Zongguang, a leading scientist at the Research Institute No. 58 of China Electronics Corporation and the rotating chairman of the Packaging and Testing Branch of the China Semiconductor Industry Association, informed China Electronics News that cracks caused by brittleness are no longer the main obstacle. He confirmed that products on glass substrates for mass production and samples for customers have successfully passed rigorous reliability tests, including thermal cycling, mechanical shock, and bending compliance.

Thus, the key competitive hurdle in the industry has shifted from solving basic brittleness issues to tackling more complex challenges related to high-precision Through Glass Via (TGV) performance, multilayer interconnects, and cost control during scaling.

Glass substrates can be applied in three chip packaging scenarios. Firstly, they serve as temporary glass carriers for already mature or near-mass-production 2D or 2.5D packages. Secondly, they can replace silicon interposers to reduce costs. Thirdly, they act as a glass substrate, replacing printed circuit boards (PCBs) or other organic substrates. In terms of TSMC's CoWoS, the second and third applications correspond to replacing the W (Wafer) layer or S (Substrate), respectively.

It is important to note that the glass substrate is intended not for the green printed circuit board in computers, but for the critical packaging layer between the chip and the printed circuit board, which determines packaging efficiency and signal integrity.

The performance of the TGV (Through Glass Via) process directly affects the overall performance of the entire board, making it the primary bottleneck. TGV accounts for over 40 percent of the processing cost of the glass substrate, with the most expensive segments being the vias themselves and metallization. The technology requires breakthroughs in three areas: forming high-precision holes using laser deep etching as the optimal path for high-density, large-size TGVs; ensuring high reliability of via metallization for uniform, dense metal deposition; and multilayer RDL interconnects and soldering, where issues such as composite substrate warping and even rupture due to the low mechanical strength of glass and thermal mismatch with multilayer wiring dielectrics dominate.

The glass substrate ecosystem in China includes substrate suppliers like Triassic and Tongge Micro, equipment manufacturers capable of local customization, and end integrators, including the Lens-Intel alliance. China Electronics Technology Corporation continues to promote standardization through its research institutes. Since the brittleness issue has been resolved, the transition to mass production now depends on improving TGV performance, the reliability of multilayer RDL, and cost curves during scaling. The strategic alliance between Intel and Lens Technology signals global endorsement of the supply chain developed in China, positioning the country for a leading role in the post-silicon substrate era.

SpaceX has scheduled the fourteenth test flight of Starship to occur before the end of August, introducing a new feature: the attempt to capture the rocket's upper stage, called Ship, using the launch tower's arms. Elon Musk communicated this detail during the company's first quarterly results meeting since its historic IPO this Tuesday (4).

Musk stated that, depending on regulatory approval, the team will attempt to capture the Ship with the tower on the next flight, which is temporarily scheduled for the end of the month. The realization of this flight is subject to obtaining the necessary authorization from regulatory bodies.

What is planned for Flight 14

The Ship constitutes the upper stage of Starship and is 52 meters high. The complete system also includes a first stage known as Super Heavy. Although SpaceX has already managed to capture the Super Heavy with the tower in three previous tests, this would be the first time such a procedure is attempted with the Ship.

In addition to the capture maneuver, Flight 14 is expected to be the pioneer in placing payloads into operational orbit, something unprecedented for Starship. In the previous flight, the 13th, conducted on July 24, it transported 20 Starlink V3 satellites, but these were placed in suborbital trajectory. If all goes as planned, according to Musk, in Flight 14 these satellites will be inserted into operational orbit.

Confidence in planning the capture was reinforced by the performance of the Ship in Flight 13. Its thermal shield withstood reentry into the Earth's atmosphere well, and the Ship landed smoothly in the Indian Ocean, near the west coast of Australia, remaining intact so far. Musk commented: 'I don't want to jinx it, but I think I would consider the thermal shield issue resolved at this point.'

Expansion Plans in AI and the Starmind Constellation

Months before the earnings call, SpaceX revealed its ambition to reach the 'total addressable market' of US$ 28.5 trillion (equivalent to R$ 146.6 trillion) annually in the long term, with almost US$ 23 trillion (R$ 118.3 trillion) belonging to the 'enterprise AI applications' segment. For comparison, the Gross Domestic Product (GDP) of the United States in 2025 was slightly less than US$ 31 trillion (R$ 159.5 trillion).

A central element of this strategy is the Starmind constellation, composed of one million artificial intelligence data centers in orbit. Musk informed during the meeting that SpaceX expects to begin launching the first operational spacecraft of Starmind in 2027. To maintain this launch pace, he projected that 'we expect the flight cadence to increase rapidly and, probably, within a year, we will be doing at least one flight per day, possibly more.'

According to the latest data from the global large language model (LLM) aggregator OpenRouter, artificial intelligence models from China have secured five of the top spots in weekly token usage.

Between July 28 and August 3, global AI models processed 56.8 trillion tokens. Chinese models accounted for 28.13 trillion tokens, significantly exceeding the 4.38 trillion tokens processed by US models. Consequently, China has maintained its lead for the fourteenth consecutive week.

Last week, nine out of the ten most popular models globally were Chinese. The DeepSeek V4-Flash model topped the rankings, utilizing 7.1 trillion tokens per week. It was followed by Xiaomi MiMo-V2.5 with 5.4 trillion tokens, Tencent Hy3 with 4.89 trillion tokens, DeepSeek V4-Pro with 3.1 trillion tokens, and Z.AI GLM-5.2 with 2.87 trillion tokens.

In terms of cost comparison, DeepSeek V4-Flash costs $0.14 per million input tokens and $0.28 per million output tokens. According to Artificial Analysis estimates, the average cost of this model is 3 cents per test, compared to 86 cents for Moonshot Kimi K3, $1.86 for OpenAI GPT-5.6 Sol, and $3.15 for Anthropic Claude Fable 5.

This comparison allows for a more realistic assessment of value because it considers the volume of data the model must process and generate to complete a task, rather than just its nominal price. A low-priced model can prove expensive if obtaining an answer requires significantly more steps.

On the Intelligence Index, DeepSeek V4-Flash scored 50 out of 100 points, matching Google Gemini 3.6 Flash, but trails by one point behind Meta Muse Spark 1.1 and Z.AI GLM-5.2. Moonshot Kimi K3 scored 57 points, while Anthropic Claude Opus 5, Fable 5, and OpenAI GPT-5.6 scored nine or more points higher.

Artificial Analysis uses a nine-parameter evaluation system covering coding, reasoning, and task-oriented capabilities to create model cards. DeepSeek's advantage in the global developer market lies in its engineering aspects: improved inference, dynamic KV caching, traffic planning, and long context optimization.

The number of foreign companies integrating the DeepSeek API is growing, as is the number of international developers using its platform. Notably, on Monday, Alibaba unveiled its largest and most powerful model to date—Qwen3.8-Max, which in size is not far behind the offering from its domestic competitor, Moonshot AI, released last month.

For millions of AI developers worldwide, this shift is becoming evident: expanded choice, reduced costs, and lowered barriers to innovation are transforming the global AI ecosystem.