February 2025 FunTech Highlights: 14 Months Later

Fourteen months ago we ran our February 2025 FunTech digest. Most of the stories we covered are now old news; the interesting part is what those headlines look like in April 2026, which trends actually shipped, and which ones collapsed spectacularly. This is the retrospective edition: what happened, what didn’t, and what the past 14 months mean for anyone building in entertainment tech right now.

How we scored the Feb 2025 stories

For each Feb 2025 headline we give a one-line verdict based on public data as of April 2026: Delivered (the thing shipped and met claims), Partially delivered (shipped but underperformed or pivoted), Flopped (failed to ship, was cancelled, or shut down). We rely on primary sources where possible — press releases, SEC filings, shutdown notices, product pages — not on commentary.

All the original Feb 2025 links are preserved at the bottom of the post so you can do your own before-and-after comparison.

CES 2025 announcements: the scorecard

CES 2025 in Las Vegas got the usual avalanche of press. Fourteen months on, here’s what actually happened:

The pattern is clear: GPU and AI infrastructure announcements delivered, while consumer hardware bets (AFEELA, ambient smart glasses) remain the hardest to ship. NVIDIA’s track record of hitting dates on high-end silicon is the cleanest signal in the industry right now.

The transparent-tech trend: peaked and retreating

We called out the Nothing Phone, Flipper Zero, and LG Signature OLED T as a visible trend in Feb 2025. Verdict: fashion-first transparent tech peaked early 2025. The Nothing Phone 4a (launched March 2026) kept the transparent back panel on the base model, but Nothing dropped transparency entirely on the Pro tier in favor of a full aluminum unibody — a telling retreat. LG’s transparent OLED remains a luxury display for hotel lobbies. Flipper Zero remains a hacker cult object, not a mass product.

Takeaway for founders: “looks cool on Twitter” and “sells 10M units a year” are different businesses. If the product-market fit is the aesthetic, the ceiling is usually Kickstarter-scale, not Fortune 500.

WhatsApp chat themes: shipped, now monetizing

Meta rolled WhatsApp chat themes and 30 default wallpapers globally in February 2025. By early 2026 it became the anchor feature for a paid WhatsApp Plus tier with 18+ premium themes. That’s a textbook playbook: ship the visible polish, then turn it into a monetization path. If you’re building a messaging product, watch this pattern closely — personalization as a paid tier is a stable 2026 revenue model. For anyone building in the communication space, the underlying WebRTC and signaling stack remains the hard part; we cover those architecture tradeoffs in our guide to video conferencing development.

Crypto: Pump.fun, meme coin rug-pulls, and the $5.5B class action

We quoted Alon Cohen calling tech altcoins “meme coins with VC exit liquidity.” His critique aged well. Within weeks of our Feb 2025 digest, the TRUMP and MELANIA meme coins rug-pulled spectacularly, wiping out retail. Pump.fun itself — the Solana-based meme-coin launchpad at the center of that wave — held the #1 spot in market share until summer 2025, when competitor LetsBonk reached 55.8% of the meme-launchpad market versus Pump.fun’s 27.4%.

Legal exposure has been the larger story. A pending $5.5B class action alleges Pump.fun operated an unlicensed gambling platform with RICO violations. The platform responded with multi-chain expansion and a native token (unlock scheduled April 12, 2026), but the compliance overhang remains. If you’re in crypto product, the Feb 2025 “VCs winning” story has become a “everyone is suing each other” story.

AI Granny Daisy: small scale, real signal

Virgin Media O2’s anti-scam AI bot “Daisy” has fielded 1,000+ calls by April 2026 and routinely keeps scammers on the line for 40 minutes with aimless chatter about knitting. O2 also uses machine learning to flag roughly a billion suspected scam/spam calls across its network in 2025–2026.

The broader lesson: AI-against-AI defense is becoming a real product category. Call-center platforms, email providers, and messaging apps are all building variants of “decoy agents” to waste attackers’ time. If you’re building in any communication category, defensive AI agents deserve a roadmap slot. For pattern-matching on the adjacent architecture, see our take on AI call-assistants APIs and our anomaly-detection surveillance piece.

Face swap and deepfakes: regulation moved faster than tech

The breezy London face-swap events we covered in Feb 2025 now sit on a very different legal foundation. The UK Crime & Policing Bill (assented February 2026) criminalized non-consensual intimate deepfakes and signaled an outright ban on “nudification” tools. The UK’s Online Safety Act treats such content as a priority offence, with platform-level duties. Across the channel, the EU AI Act’s deepfake transparency rules come into force August 2, 2026: any AI-generated or manipulated image, audio, or video must be clearly labeled, with few exceptions.

For product teams this is the single biggest content-compliance shift of the last 14 months. If you ship any feature that lets users apply a face or voice to another person’s media, you need a consent layer, a watermarking layer, and a moderation path. “Fun” is no longer a product positioning; it’s a liability category. Our emotion recognition in video conferencing guide walks through the adjacent consent-UX mechanics that transfer cleanly here.

AI video generation: Sora dies, Veo and Runway inherit the market

This is the post-Feb-2025 story that changed the creator-tool map most. OpenAI’s Sora went from flagship hype to shutdown in under 18 months. The company confirmed in March 2026 that the consumer Sora app would cease operation April 26, 2026, with the API following September 24, 2026. The cited reasons: roughly $1M per day in inference costs, user base collapse from over 1M to under 500K, pending copyright litigation, and competitive pressure.

The inheritors of the market, as of April 2026:

Google Veo 3.1 Lite, released March 31, 2026 and made free to every Google account holder on April 2 — effectively to billions of users. Native audio, extended duration, improved character consistency.

Runway Gen-4.5, benchmarking as the top-ranked video model in independent leaderboards, particularly strong on photorealism and motion control.

Kuaishou Kling 3.0+, a credible Chinese competitor with comparable fidelity and a different training-data tradeoff on licensing. If you’re building in the streaming or VOD space, the pragmatic pipeline choice in April 2026 is a Veo-first or Runway-first API with a self-hosted fallback for cost-sensitive workflows. Our overview of the best AI video enhancement tools in 2026 breaks down the integration surface.

Agentic AI: the category that didn’t exist a year ago

At Feb 2025 publication, “AI agents” meant research demos. Fourteen months later, agentic AI is the defining consumer-product category of the year. Anthropic’s Claude Computer Use shipped in Claude Pro/Max, Claude Code, and Cowork desktop (including the product you’re reading this in) in March 2026. OpenAI’s Codex Computer Use followed in April, bringing background Mac automation with 90+ plugins and persistent memory.

The underlying glue layer is the Model Context Protocol (MCP), Anthropic’s open standard for connecting models to tools and data. By March 2026 MCP had 97 million installations. The Linux Foundation formed the Agentic AI Foundation in December 2025; OpenAI’s AGENTS.md, Block’s goose, and most agent frameworks now interoperate through the protocol.

Translation for product teams: in 2026, shipping a feature that does not expose its core actions to a user’s AI agent of choice is a silent retention loss. Every meaningful SaaS product is now building an MCP server or an agentic integration layer alongside traditional UI. We documented a similar pattern for iOS in our recent 2026 iOS development retrospective: App Intents are the iOS equivalent of MCP, the same shape of bet.

Humanoid robots: from demo reel to paid pilots

Figure AI’s Figure 03 is in paid pilot production at BMW’s Spartanburg plant for material-handling. Figure publicly targets around 90,000 units per year at full ramp; as of April 2026 production is still measured in hundreds per month, not thousands. Apptronik and Unitree are shipping smaller pilot fleets. The category has clearly exited lab-demo status, but the long tail of reliability, maintenance, and cost-per-hour economics is very much unsolved. For anyone in industrial software, the question is no longer “will humanoids matter?” but “how soon do we need fleet-management APIs for them?”

Spatial computing: Apple’s quiet year, Meta’s Hyperscape

Apple Vision Pro Gen 2 has still not shipped. A mid-cycle M5 refresh with a new knit band arrived October 2025 at $3,499, but a true successor is at least into late 2026 territory. visionOS 26 brought spatial AI scene authoring, improved Personas, and PS VR2 controller support, but Vision Pro adoption in 2026 remains modest; it’s the “developer platform” Apple originally pitched, not the mainstream device.

Meta’s Hyperscape Capture, by contrast, is quietly one of the most interesting consumer experiences of 2026. Scan a physical space for 5–10 minutes with a Quest 3 or 3S, wait 1–8 hours for Gaussian-Splatting processing, and get a photorealistic VR digital twin you can invite friends into. Multiplayer support arrived in April 2026. For enterprise verticals — real estate, construction, training — this is a cheaper spatial-capture pipeline than anything Apple offers.

If you’re exploring spatial software, our Vision Pro business-case article and our AR/VR development page are the right starting points.

AI wearables: Humane and Rabbit became cautionary tales

The Humane AI Pin was shut down on February 28, 2025 — 18 days after our original digest went live. HP acquired the assets for $116M and folded Humane’s engineering team into HP IQ. Post-shutdown Pins are bricks; HP offered refunds for units bought after November 2024.

Rabbit R1 sold around 100,000 units and drowned in returns. RabbitOS 2, promised in late 2024, finally shipped September 2025 — more than a year late. Early-2026 reports describe unpaid salaries at the company. The lesson is consistent: shipping hardware before the software is ready is fatal. Building a credible hardware-plus-AI product in 2026 means treating the cloud and on-device model story as table stakes, not marketing.

AI companion apps: survived, but under fire

Replika remains operational despite a €5M Italian GDPR fine. Character.AI, with 233 million registered users as of Q1 2026, faces three concurrent lawsuits (filed September 2025) alleging that companion-style behavior contributed to teen self-harm. The company rolled out a safety overhaul late 2025. The narrative around AI companions has shifted from “endearing” to “regulated-harm category.” Product builders in dating, social, or mental-health adjacent categories should now treat companion-style features as high-scrutiny design territory — closer to a medical-device framing than a consumer toy.

AI music: licensing replaces free training

Warner Music Group settled with Udio and Suno in November 2025, folding into licensing deals and “licensed models” that Suno is rolling out in 2026. Sony Music has not settled as of April 2026; that case continues. Independent class actions (Justice v. Suno in June 2025; Woulard in October 2025) are pressing the same claims. A January 2026 joint suit by UMG, Concord, and ABKCO alleges $3B+ in damages over 20,000 songs. The dynamic mirrors what happened in image generation two years earlier: legal pressure forces licensing; licensed-model tiers replace the free-training model. If you build audio or video tools, your training-data pipeline needs a licensing story before it needs a model story.

Stargate: not vaporware

The OpenAI-led $500B Stargate infrastructure project has visible progress. The Abilene, Texas flagship site has two buildings operational since September 2025, with six more expected by mid-2026 and 450,000+ NVIDIA GB200 GPUs targeted. Additional US sites are under active construction in Shackelford TX, Doña Ana NM, Lordstown OH, Milam TX, the Midwest, and Saline MI. International projects are underway in the UAE, Norway, the UK, and Argentina. Roughly $400B of the $500B commitment is in play. By 2029, Stargate is projected to draw over 9 gigawatts — a load comparable to New York City. Whatever your view of the economics, this is real capital deploying real hardware.

Generative game content: shipping, but still niche

NVIDIA ACE smart NPCs shipped in inZOI (March 2025) and received a PUBG Ally long-term-memory update in H1 2026. Total War: Pharaoh showed a public ACE tech demo at CES 2026. Google DeepMind’s Genie 3 and the Oasis class of interactive generative models remain research frontiers, not shipping products. If you’re building in gaming, generative NPCs are “interesting to prototype, too expensive to ship everywhere” in April 2026. That will change, but the inflection is still at least one more engine cycle away.

Creative coding: subpixel Snake was a one-off

Patrick Gillespie’s subpixel Snake demo never spawned a movement. Creative-coding gems are, by nature, individual artistic statements more than trend starters. The healthier signal came from the demoscene and shader-art communities, which had a robust 2025 without making the mainstream tech press. If you enjoy that world, the annual Revision competition continues to be the best yardstick.

What all of this means if you’re building right now

Six product-team implications from the 14-month retrospective:

1. Agentic AI surface is now table stakes. Expose your core actions over MCP or an equivalent agent protocol. iOS, macOS, and Windows agent integrations all rely on this pattern; so does Claude Computer Use. Not optional past Q2 2026.

2. Video generation economics are a compute problem. Sora shut down because inference costs outran revenue. If your product relies on frontier video generation, model a conservative gross-margin case before you commit. Consider hybrid pipelines (cheaper classic video + selective generative inserts).

3. Deepfake features are a compliance category, not a marketing angle. UK criminal law, EU AI Act Article 50, and state-level US biometric statutes all treat synthetic-media features as regulated surfaces. Your consent and watermarking design is now a first-class product concern.

4. Hardware without shipped software dies fast. Humane and Rabbit are the 2025 tombstones. If you are building a device, budget twice the software runway and hire model-ops talent before you hire mechanical.

5. Licensing precedes models in media. In 2024 the fight was “can we train?” In 2026 it is “who have you licensed from?” Expect audits; prepare the paper trail.

6. Infrastructure spend is real. Stargate, Amazon Trainium, Microsoft Atlantis — the 2025–2026 data-center capex cycle is as large as anything the industry has ever seen. Pricing on inference will drop again in 2026 H2; model your roadmap against cheaper, faster frontier models by end of year.

What Fora Soft actually shipped in this window

Between Feb 2025 and April 2026 we delivered AI-integration work on platforms including ProVideoMeeting (AI summarization and action extraction on top of an existing WebRTC stack), CirrusMed (HIPAA-grade telemedicine with patient-engagement signals), Meetric (AI sales-video platform), and Mindwibe (AI-powered iOS dating app integrating Image Playground and Writing Tools). The consistent pattern: WebRTC or native media pipeline as the foundation, agentic AI layer on top, consent and compliance UX from day one.

Mini case: agentic summarization for a conferencing client

A late-2025 engagement: an enterprise conferencing client asked us to replace their basic transcript summarizer with an agentic layer that could file action items directly into Jira, Asana, or Linear. We used Anthropic’s Model Context Protocol for tool binding, Claude 4.5 Sonnet for orchestration, and a small classifier to decide when to call external tools. Total build was 640 senior-dev hours, fully loaded cost around $96k, shipped in eight weeks. Adoption among pilot customers reached 78% weekly-active inside a month. The piece that changed the trajectory was not the model choice but the tool-binding protocol; MCP made the integration reusable across every downstream tool the customer’s team used.

What a Feb 2026 digest would look like next year

If you’re reading this as a Q2-2026 product builder, here’s the bet set for next year’s retrospective: agentic AI becomes invisible infrastructure (like REST APIs did in 2013); the first humanoid-robot consumer lawsuit lands; at least one major AI-music platform goes IPO or acquires its way to exits; the EU AI Act’s Article 50 transparency rules force deepfake watermarking to become a browser-level standard; and at least one frontier-model vendor announces a significant price cut on video generation as compute supply catches up. We’ll revisit these predictions twelve months from now.

FAQ

The one-sentence retrospective

Consumer hardware bets fell hard (AFEELA, Humane, Rabbit, Sora-as-app), agentic AI emerged as 2025’s defining software category, and entertainment tech entered a regulated era — if you’re shipping in this space, plan your roadmap around the new rules rather than the old hype cycle.

Read next