AI Competition Radar: September 13–20, 2026
A scout-focused weekly roundup of Tencent's four hackathon first prizes, Zhihu's six campus-built apps, MIT delta v's 13 Demo Day startups, and Blackbaud's eleven-company cohort, plus three announced results still missing a roster.
Tencent Cloud's second global hackathon, Zhihu's campus hackathon in Beijing, MIT's delta v Demo Day, and Blackbaud's social-impact startup showcase all named their winning teams inside the September 13–20, 2026 Pacific-time window this issue covers. Three further events — the AI Infra Summit Hackathon in Santa Clara, Korea's LG Aimers, and a Movistar-backed student competition in Peru — confirmed that a result exists and published no team names; this issue holds them in a separate unresolved queue.
Two things stand out across the named results. At Tencent's final and again at MIT's Demo Day, the winning work was software that other people can pick up and run, and several of those teams say they have already put it in front of institutional users. The rest of this issue is what each result actually is, and the question a scout would want answered before sending the first email.
Fast scan
| Event and organizer | Result signal | Project or team | Public evidence | First diligence question |
|---|---|---|---|---|
| Tencent Cloud Hackathon 2026 global finals, Shenzhen (result announced September 13) | Four first prizes across the agent, game, and animation tracks; 10,122 participants, 1,000+ submissions, 508 universities, 42 finalist teams | 星光漫界, ZNL, Anchor (HKUST), 《Autism is POWER》 | Organizer recap names all four first-prize teams and their entries 1; the animation winner is also reported by name 2 | Which of the four teams has a user outside the competition, on a toolchain that is not Tencent's? |
| Zhihu Campus Rising Season hackathon, Beijing 798 (concluded September 19) | Six prizes and three category awards named; 1,878 developers from 450 universities, 435 entries, 15 finalists | MeetMind 学脉 (Tsinghua), 看山有碗汤, 知乎·回响, 问山 (ThreadPeak), 知乎脑洞游乐园, 心动小屋 | Organizer account of the final lists every prize and describes each build 3 | Which of these six retains users who never entered the competition? |
| MIT delta v Demo Day, Cambridge, MA (Demo Day held September 11; account published September 16) | 13 teams and 28 founders presented; each received $75,000 equity-free; selected from more than 200 applicants | Neural Physics, Cerebrus AI, Talys, and ten more | MIT News account names the teams, the funding terms, and the founders' claims 4 | Is Talys's $325 million of processed client spend contracted volume or a pilot? |
| Blackbaud Social Good Startup Program showcase, Charleston, SC (announcement published September 18) | 11 cohort companies pitched; Judges' Choice and People's Choice both announced October 1 | DonorAtlas, Givalgo, Kindora, PlanPerfect, Vitalcy, and six more | Program announcement names the full 2026 cohort and the voting timeline 5 | Which of the eleven has revenue that does not depend on Blackbaud's marketplace? |
Confirmed results
Tencent Cloud Hackathon: four first prizes, three tracks, one unpublished build
Tencent Cloud closed the second edition of its global hackathon on September 13 at its Shenzhen headquarters. The contest ran 150 days across five regions, drew 10,122 participants and more than 1,000 submissions from 508 universities, and sent 42 teams to the final across three tracks — agents, games, and animation. 1
First prizes. In the agent track, 星光漫界 took the top prize with 《星光漫界智创平台》 and ZNL took a shared top prize with 《Psych-MAP》. In the game track, Anchor, a student team from the Hong Kong University of Science and Technology, won with 《锚点》 (Anchor), a first-person narrative puzzle game that alternates between a mother living with Alzheimer's disease and the daughter who cares for her. In the animation track, a Hong Kong secondary-school team won with 《Autism is POWER》, a short film about an autistic protagonist built with AI techniques for character and scene consistency. 12
The organizer's recap names the two agent-track winners and their project titles and publishes no description of either build. That leaves the agent-track podium, the most scout-relevant one, as a title and a team name until the teams publish something themselves.
The second prize is the one already in production. A team led by Professor Zeng Hui of Guangzhou Panyu Polytechnic took second prize in the agent track with 普惠金融数字员工 (inclusive-finance digital employees). The competition set a seven-day build window, and the team delivered nine digital employees for a private fund that manages 20 stocks, then put them into trial on the firm's real business. 6
A named regional winner. 小福狸, which took first place in the East China region, was exhibited at the Shenzhen final. Its product lead, 杨夏鸣, describes it as a work-memory layer for lawyers and other knowledge workers: a persistent desktop entry point that captures, with low interference, how a person and an agent complete a task together, then assembles the finished task into a reusable PlayBook. The system also extracts a user's working style and habits into a portable profile document that another person can import. 7
Where the entries came from. Roughly 30% of the participants work in front-line roles in catering, law, design, and foreign trade, and more than 30% of the new game and animation track entrants are not programming majors. The builds ran on three of Tencent's own agent platforms — WorkBuddy for workflow construction, CodeBuddy for the design-to-deployment path, and Miora for image, video, and 3D generation. 7
That last detail is the diligence question for all four winners: the workbench they built leans on the sponsor's toolchain, so a scout wants to know which parts of it run anywhere else.
Zhihu Campus Rising Season: six apps built on a platform's own data
Zhihu closed its campus hackathon on September 19 in Beijing's 798 art district. The programme ran 35 days of campus talks and a 48-hour build sprint, drew 1,878 developers from 450 universities, and received 435 entries; 15 teams reached the final. Entrants worked on three tracks and could call Zhihu's hot list, its on-site and web search, its direct-answer agent, and, after OAuth authorisation, a user's creation and bookmark data. 3
First prize — MeetMind 学脉 (Tsinghua University). The project turns a user's saved Zhihu bookmarks into a one-on-one tutor that narrates and draws at the same time. A learner can interrupt when the explanation loses them, then take a quiz, work through flashcards, or explain the concept back; concepts that still fail send the learner to two articles the system picks from Zhihu. 3
Second prize — 看山有碗汤 (Renmin University of China, Nanyang Technological University and others). The team automated the "lateral thinking puzzle", a guessing game that normally needs a human host. A model sets the puzzle and judges up to six yes/no questions per player, a program handles question counting and clue merging, and the player assembles the answer. Each round closes with an analysis of the player's questioning pattern. 3
Second prize — 知乎·回响 (Peking University, Fudan University, University College London). The project extracts three to six further-explorable insights from a long answer and keeps each one's original position in the text. A reader picks the position closest to their own, and the system sorts related content into aligned, opposing, and undecided, with the sorting basis stated. 3
Third prize — 问山 (ThreadPeak). The system interviews a learner about what they want to do, their current level, their deadline, and their resource limits, then converges on a branching learning route on a single canvas. When a question needs another person's experience, it routes to an author through Zhihu's contributor network. 3
Third prize — 知乎脑洞游乐园 (China Agricultural University). The player sits in a decision seat and debates four AI courtiers holding different positions, choosing each round between a steady, aggressive, bartering, or breakthrough move. At the end the system writes the whole run up as a Zhihu-style long post that can be published back under the original question. 3
Third prize — 心动小屋 (Wenzhou-Kean University). The project combines story reading, character collection, and multi-character AI interaction into a single show: a user assembles three guests from open story fragments, then runs a 20-day dating-show schedule in which dialogue, joint tasks, anonymous messages, and key choices change the relationships. The model performs the plot live against the current progress. 3
Category awards went to 知乎·回响 for social innovation, MeetMind 学脉 for learning tools, and 看山有碗汤 for game creativity. Zhihu also put numbers on the platform these apps sit on: its AI Works gallery holds more than 2,600 AI projects, and more than 17,600 developers have connected to its data platform. 3
Every one of these six is a consumer product on a single platform's content and social graph. The first question for a scout is what each one looks like with that platform removed.
MIT delta v Demo Day: 13 teams, $75,000 each, and a health-care agent already in the field
MIT's delta v accelerator held its 15th Demo Day on September 11 in Kresge Auditorium, and MIT News published its account of the cohort on September 16. Thirteen teams presented two-minute pitches, drawn from more than 200 applicants, twice the 2025 number. Each team received $75,000 in equity-free funding, a sharp increase on earlier years, and the cohort ran to 28 founders across undergraduates, graduate students, and postdocs. The Trust Center also announced a partner model of more than 125 founders from companies including HubSpot, Okta, and Kayak, alongside industry experts and early-stage investors. 4
Three teams made claims specific enough to diligence.
Neural Physics is building foundation models for physics, aimed at compressing the design and validation cycles that manufacturers run through simulation. Co-founder and PhD candidate Mohamed Elrefaie put the cost argument in his pitch: the U.S. advanced manufacturing sector loses roughly $245 billion a year to engineering delays, and most of that time goes into running simulations and making engineering decisions. 4
Cerebrus AI monitors deployed AI agents for behaviour that departs from an approved baseline. The system builds a behavioural profile for each agent and flags activity that falls outside it. Co-founder and MBA student Griffin Potrock framed the product around three executive questions: where the agents are, what they are doing, and what they have access to. 4
Talys uses AI agents to find savings across pharmacy, operations, and third-party spend for health-care organisations. Co-founder and MBA student Nicolas Berzin said the company is already working with health systems and has processed $325 million in spending. That figure is the company's own, given in a pitch, and the published account records it as such. 4
The remaining ten teams, most of them hardware or deep-tech, were Alpaca (hardware and software for individuals to host frontier models), Banzai (an agent that diagnoses home repairs), Bizon Labs (a platform for engineering lipid nanoparticles), Cortheon (AI design optimisation for foundries making complex metal parts), Exo AI (back-office automation for financial institutions), Pixology (agentic AI for media-rights sales materials), Robox (a design engine for physical automation), The Trade Lab (tariff and supply-chain navigation for importers), RBT Resources (a portable blood-transfusion device), and Gander Robotics (a hand-thrown autonomous drone submarine for man-overboard rescue). 4
Blackbaud: eleven companies pitched, and both winners wait until October 1
Blackbaud (NASDAQ: BLKB) announced on September 18 that it had hosted the annual showcase for its Social Good Startup Program. Eleven companies in the 2026 cohort pitched live, and a panel of Blackbaud judges questioned each founder; the public vote runs alongside the judging. Both the Judges' Choice and the People's Choice winners are announced on the main stage at bbcon 2026 on October 1, and voting closes at 11:59 p.m. ET on September 30. The prize is an exhibitor-level sponsorship at bbcon 2027 plus marketing and executive coaching. 5
The cohort is AppsNXT, Bridgit, DonorAtlas, Givalgo, Grasshopper Signup, Kindora, MatchNice, NetOutlook, PlanPerfect, Untapped Solutions, and Vitalcy. Five of the eleven describe themselves in AI terms: DonorAtlas as an AI-native donor-intelligence platform, Givalgo as AI agents that automate grantmaking and due diligence, Kindora as AI-powered grant discovery, PlanPerfect as an AI-assisted strategic-planning and risk platform, and Vitalcy as AI-powered engagement for nonprofit communities. 5
Two signals came with the announcement. One nonprofit raised $4.5 million after implementing the strategic-planning platform in the cohort, and Blackbaud says it has supported more than 95 startups across 12 cohorts since 2020, with 80 still operating. Both figures come from Blackbaud. 5
Result announced, roster unresolved
AI Infra Summit Hackathon (Kisaco Research and lablab.ai)
The official hackathon of AI Infra Summit closed its online build on September 16 and its on-site round at the Santa Clara Convention Center on September 17, with 3,018 participants, 617 teams, and 163 projects across four tracks: Intel online (103 submissions), Qualcomm on-site (11), SiMa.ai on-site (12), and Intel on-site (8). 89
The results page states that judges are reviewing every finalist and that results will be published there once judging concludes. The public leaderboard currently reflects community hearts, where Velos Space (Team Relogic) and OpsPilot AI (Quantum Coders) lead; community standing is not a judging result, and this issue treats the event as unresolved until names appear. 8
LG Aimers (LG, South Korea)
LG ran the country's largest AI hackathon, drawing 3,400 young participants, with 96 advancing to the championship round through competition of 34 to 1. The top three teams shared 10 million won, and winners can skip LG's document-screening stage when they apply. More than 23,000 people have taken part since the programme began in 2022, and a recruitment fair ran alongside the final with LG AI Research, LG Electronics, LG Uplus, and LG CNS. 10
The published account names no winning team and no project. For a scout the event reads as a hiring funnel rather than a deal source until the roster appears.
Hackathon AI Telecom Challenge (Movistar Perú, Universidad de Lima, Innova Ulima)
More than 850 university students registered, 90 teams submitted proposals, 20 teams reached the final with 11 Movistar mentors attached, and three teams were selected as winners. Partners included NTT DATA, Google for Education, Lenovo, and Cisco. Women made up 32% of the members of the 20 finalist teams. 11
The competition organiser published the outcome and no team names, project titles, or demo links. The reference above is a news report, and it carries the same gap.
September 21–October 15 watchpoints
- September 25: Hack the North 2026 at the University of Waterloo announces its winners. The event drew 1,045 participants, and judges weigh originality, user experience, technical complexity, and a live demo. 12
- September 30 to October 1: Public voting for Blackbaud's Social Good Startup Program closes on September 30, and both winners are announced at bbcon 2026 on October 1. 5
- October 13–15, San Francisco: The 200 finalists of TechCrunch's Startup Battlefield 200 pitch at Disrupt, which runs at Moscone West. TechCrunch named a further five venture investors to the judging panel on September 16. 1314
- On the AI Infra Summit Hackathon: the results page at lablab.ai is the place the organiser says it will publish judging results. 8
- Still unpublished: DevNetwork's winner announcement for the [API + Cloud + AI] Hackathon still directs readers to the general project gallery rather than a ranked table, roughly two and a half weeks after the update went up. 15
- New this week: Tencent Cloud opened a campus agent challenge with Chongqing University on September 18, running to November 21, which suggests the WorkBuddy-led format from the Shenzhen final will produce another roster in November. 16
Scout priorities
- Screen the two agent-track winners before anything else. 星光漫界 and ZNL hold the only agent-track first prizes from Tencent's final, and neither build has a published description. Ask each team for a two-minute demo recording and the parts of their stack that come from outside WorkBuddy, CodeBuddy, and Miora. 1
- Talk to Professor Zeng Hui's team at Guangzhou Panyu Polytechnic. Nine digital employees built in seven days for a fund managing 20 stocks, already in trial on live business, is the most concrete deployment claim out of Shenzhen this week. Ask what the trial is measured on and who signs off on an agent's output. 6
- Diligence Talys on the $325 million. The figure is the team's own, stated in a pitch. Ask whether it is contracted spend under management, pilot volume, or gross claims processed, and get the number of health systems behind it. 4
- Hold the Blackbaud eleven until October 1, then move on the two winners. The cohort is an accelerator pipeline rather than a podium, and the judges' and public results are three days apart from each other on the calendar. 5
- Treat the three unresolved events as watch items, not leads. The AI Infra Summit Hackathon, LG Aimers, and the AI Telecom Challenge each confirmed a result without a name; checking their own pages in the next two weeks is cheaper than trying to reconstruct a roster from coverage. 81011
This roundup runs weekly; the next issue covers September 20–27.
References
- 12026腾讯云黑客松总决赛落幕 见证 Agent 把创作权交给每个人
dutenews.com
- 2香港高一学生拿下腾讯云黑客松决赛一等奖
static.nfnews.com
- 3知乎黑客松校园新锐季收官：450所高校1878位学生开发者参赛
news.sohu.com
- 4
- 5
- 6教师获腾讯云黑客松AI智能体全球总决赛二等奖
gzpyp.edu.cn
- 7不懂编程也能「手搓」AI产品！腾讯这场大赛，揭开了OPC创业门槛的转移
news.qq.com
- 8
- 9AI Infra Summit Hackathon
lablab.ai
- 10
- 11
- 12Hack the North 2026: Canada's Biggest Hackathon
hackthenorth2026.devpost.com
- 13
- 14TechCrunch Disrupt 2026
techcrunch.com
- 15And the winner is...
api-cloud-ai-hackathon-2026.devpost.com
- 16腾讯云黑客松官网
tch.cloud.tencent.com
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.