Tencent Raises Capex 176 Percent For AI First Pivot Amid DeepSeek And Grok Launches
English version above, Chinese version below.
Trend One: Frontier models clash on the same day, ending single-vendor dominance
On August 13, two frontier models launched within 24 hours of each other. DeepSeek released the stable version of V4 Pro, updated on its site under the model id DeepSeek-V4-Pro-0813 and made available via API, with Agent capabilities materially strengthened through Responses API and Codex integration. On several benchmarks the model approaches or clears Claude Fable 5, while holding a price-to-performance lead. SpaceXAI shipped Grok 4.6 the same day, building on Grok 4.5 to improve long-running agent tasks, complex interactions and vision. Grok 4.6 posted top scores on InferenceEval (46.9%), OfficeQA Pro (63.2%) and the Harvey legal-agent benchmark (22.0%), and tied GPT-5.6 Sol on the Artificial Analysis Intelligence Index, a nine-benchmark composite.
The same-day timing is not coincidental background. On August 13, Reuters reported that Google co-founder Sergey Brin has spent recent months pressing DeepMind staff to commit fully to Gemini after Alphabet fell behind Anthropic and OpenAI. In April Brin addressed an all-hands of several hundred employees urging the lab to move faster. Internal testing showed the new flagship Gemini still trailing on coding, so Google pushed its release back by two months in August. On August 5, DeepMind replaced its leadership — Demis Hassabis stepped down as CEO into a chairmanship role, with deputy Koray Kavukcuoglu taking over, while two original Gemini technical leads left to found a startup. Anthropic and OpenAI are now the reference points that both a Chinese lab and Google's own internal efforts are measured against. (Source: ithome 989/023, 989/035; cb_doge on X)
Trend Two: Tencent's WeChat turns "AI-first", and the compute supply chain follows
On August 13, Tencent's WeChat team published its self-built large language model WeLM on X. Two configurations are disclosed: WeLM-80B (80B total parameters, 3B active) already deployed inside WeChat's native AI assistant "Xiaowei", handling chat and search, calling WeChat-native functions and mini-programs; and WeLM-617B (617B total, 23B active), a mixture-of-experts model still in development, targeting complex in-ecosystem tasks like mini-program development and tool generation. Separately, Tencent's Hunyuan model completed a rapid preview-to-stable iteration this quarter — Hy3, a 295B-total / 21B-active MoE model with 256K context, saw call volume grow over 68x within a week of launch, with a larger Hy4 soon to follow.
The strategy was spelled out the prior evening. On August 12, Tencent president Martin Lau told the Q2 earnings call that WeChat will become an "AI-first" ecosystem — a user gives one instruction and the system executes. Lau framed it against QQ's PC-era role and WeChat's 10x amplification in mobile, arguing the ecosystem can convert its existing monetization into "tremendous value." The numbers behind the pivot: Q2 revenue of 204.79 billion yuan, up 11% year over year; capex of 52.78 billion yuan, up 176%, with GPU supply expected to loosen by late this year into early next.
That capex flows downstream. On August 12, Hon Hai (Foxconn) reported its Q2 structure shift: cloud and networking products crossed 50% of quarterly revenue for the first time at 51%, while smart consumer electronics including iPhone fell to 29%. AI server revenue also passed 50% of the total. Q2 net profit rose 35% year over year to 59.97 billion New Taiwan dollars. Foxconn is preparing volume production for Nvidia's next-generation Vera Rubin server racks, targeting Q3 production-readiness and Q4 shipments, with Vera Rubin expected to become a significant product in 2027. Rotating CEO Chiang Jih-heng called the cloud growth structural rather than cyclical. (Source: ithome 989/007, 988/977, 988/966)
Trend Three: Governance and consent become the new battleground
Three governance signals landed in the same window. Bloomberg-linked reporting on X indicated the White House is weighing whether to bring open-source AI models under its pre-release safety-testing framework, which currently covers only closed frontier labs like OpenAI and Anthropic. If open models reach equivalent capability, they could face up to 30 days of pre-release testing — and officials are weighing the risk that this slows open-source development in the US.
On August 12, two consent disputes surfaced. Anthropic began inserting an invisible watermark into Claude's output to satisfy the EU AI Act's transparency code, which requires AI-generated content to be machine-identifiable. Some Reddit users complained the watermark would expose them using Claude for jobs or coursework; the majority of commenters backed the measure as a way to trace algorithmically produced text. The same day, Twitch announced it would train Amazon's generative models on streamers' content by default unless creators opt out — an opt-out stored in channel security settings, not an opt-in. Twitch's chief product officer told a live audience of roughly 3,000 that an opt-in design would draw "nobody," an acknowledgment that triggered hundreds of anti-AI messages in chat.
A fourth incident showed how fragile the information layer itself can be. On August 12 (local time), Google Search briefly displayed a death notice for OpenAI CEO Sam Altman after his Wikipedia page was vandalized — edited 22 times in a single day, including a false claim he was assassinated. Google corrected the result within hours and said it was not a manual change; the knowledge panel draws on multiple public sources, of which Wikipedia is one. (Source: kimmonismus on X; TechCrunch 2026/08/12; ithome 989/024)