Solid Numbers, but a Troubling Signal

On paper, Reddit just delivered a quarter most companies would envy. Revenue hit $805 million — a 61% jump from last year. Net income surged 183% to $253 million. Both figures sailed past what Wall Street analysts had predicted. The company even guided for next quarter revenue between $860 million and $870 million, with healthy pre-tax earnings.

So why did the stock tumble more than 10% in after-hours trading?

The answer lies in a single word CEO Steve Huffman used in a letter to shareholders: choppy.

The ‘Choppy’ Search Referral Problem

Huffman wrote that search referrals — the traffic Reddit gets from people clicking links on Google and other engines — were “choppy in the quarter” and became “more volatile” toward the end of Q2. “The bigger picture is unchanged: the commercial business is strong,” he added.

Investors didn’t buy the reassurance. The drop suggests they see the choppiness as a symptom of something bigger: Google’s AI Overviews and other generative search features are starting to eat into Reddit’s organic traffic.

It’s a delicate moment. In 2024, Reddit signed a data-licensing deal with Google, letting the search giant train its AI models on Reddit content. But now Google’s own AI summaries — which pull answers directly from sources without requiring a click — appear to be siphoning off the very users Reddit relies on for ad revenue and engagement.

Reddit has already signaled uncertainty about renewing that partnership.

U.S. Users Dipped — and Analysts Pounced

Reddit’s global daily active unique visitors grew, but the U.S. number slipped slightly: from 53.5 million in Q1 to 53.2 million in Q2. That’s a tiny drop — 0.6% — but it set off alarms.

During Thursday’s earnings call, one analyst didn’t mince words. “I don’t want to belabor this point, but your stock is down sharply because there is just a sense from investors that you have a — I don’t want to, to be blunt — a user problem, especially in the U.S.,” the analyst said.

The same analyst laid out the fear plainly: logged-out traffic is under pressure as search shifts to AI, cutting off referral flows. And if fewer people discover Reddit through search, fewer convert to logged-in, registered users. That’s the core worry.

The analyst then asked directly: “Do you see any world where you’re not licensing data to Google and OpenAI next year?”

Huffman Bets on Human Connection

Huffman pushed back, arguing that Reddit’s value isn’t in raw search traffic — it’s in authentic community conversation. “Reddit is communities and conversation. Communities are universal, and so we think we have in the U.S. content for everyone, and it’s a matter of revealing that,” he said.

On the Google relationship, Huffman was more cagey. He noted that Reddit’s ties with Google predate any formal licensing deal. “I don’t think there’s a binary outcome,” he said. “We will make sure that we’re maximizing the value for Reddit.”

That vagueness didn’t calm the market. Investors want to know whether Reddit can keep growing its user base in an AI-dominated search world — or whether it’s becoming a raw material supplier for the very tools that threaten its traffic.

What This Means for Reddit’s Future

The Q2 numbers are objectively strong. Revenue growth at 61% and net income nearly tripling are rare for a company that only went public in March 2024. But the market is forward-looking, and the forward picture is murky.

Three questions hang over Reddit:

• Can Reddit diversify traffic sources beyond Google search, especially as AI summaries reduce click-through rates?
• Will the U.S. user decline turn into a trend, or was it just noise in a strong quarter?
• What happens when the Google data-licensing deal comes up for renewal? If Reddit walks away, it loses a revenue stream. If it stays, it may be feeding the beast that eats its traffic.

For now, Reddit has the financial cushion to experiment. The core product — real conversations among real people — remains differentiated from AI-generated content. But the stock’s after-hours slide is a clear message: Wall Street is watching how Reddit navigates a search landscape that’s shifting faster than anyone expected.

Meta’s enterprise AI ambitions go far beyond customer service bots

Mark Zuckerberg has a message for investors: don’t think of Meta enterprise AI as just another chatbot play. During the company’s second-quarter earnings call on Wednesday, the Meta CEO laid out a far bigger vision — one that spans APIs, direct compute sales, and even internal productivity tools repackaged for external customers.

In June, Meta quietly entered the enterprise AI market with an AI agent designed to help businesses handle customer service, support, and daily operations. But that was just the opening act. Zuckerberg made it clear that the company sees a much larger opportunity ahead.

“We see a large enterprise opportunity to sell to businesses, including APIs, business agents, potentially selling compute directly, and other services that we’re building for large customers,” he said.

These additions could help Meta build new revenue streams beyond its core advertising business — which still drives the overwhelming majority of its revenue — and subscriptions, which contribute a much smaller slice.

Advertisers first, then the rest

For now, Meta is focused on serving its existing base of millions of advertisers. The company plans to offer AI agents that work across its messaging apps and other platforms, allowing businesses to interact with customers through an AI interface.

“And, just like the ad system, effectively, we will get paid when we deliver results for those businesses,” Zuckerberg explained. “We view this as an extension of the sales and the partnerships that we have with many millions of advertisers and hundreds of millions of small businesses that use our platforms.”

That’s a smart starting point. Meta already has deep relationships with small and medium businesses through its ad platform. Offering AI agents as a natural extension of that relationship feels less like a pivot and more like an upgrade.

Internal tools, external customers

But Zuckerberg also hinted at a bigger play. He described how Meta could eventually sell its own internal tools — the software the company built for its own engineers and product teams — to external customers.

“We’re building coding and developing and internal productivity tools partially because we need to build them ourselves, and we need to make sure that we have tools that are tuned for ourselves,” he said. “Now that we have those, we feel like there’s a large opportunity to serve — whether that’s small businesses or larger businesses.”

It’s a classic move: build something for yourself, then sell it to the world. But Zuckerberg admitted this won’t be easy. Selling to enterprise customers, he acknowledged, is a “different muscle” than the one Meta has historically flexed.

Meta’s compute play: sell now or save for later?

One of the more interesting parts of the call was Zuckerberg’s discussion of compute sales. Meta has invested heavily in AI infrastructure, and the company sees an opportunity to sell some of that compute capacity to enterprise customers — at a premium.

The company pointed out multiple times that it currently has the opportunity to sell compute at “a significant premium over what we paid for it.”

But Zuckerberg cautioned against cashing out too quickly. “It would be foolish,” he said, to “sell all of the compute and take a short-term profit.” Instead, he described Meta’s approach as a “portfolio” that balances long-term and short-term plans for its compute infrastructure.

“As we get closer to personal superintelligence, we are … going to need hardware that allows you to seamlessly interact with it,” he noted.

That’s a reminder that Meta’s enterprise AI ambitions are tied to a much bigger bet on agentic AI — systems that can act on a person’s or business’s behalf, rather than just answer questions.

Consumer AI agents and smart glasses are also on the table

Enterprise customers aren’t the only ones getting AI agents. Zuckerberg also promised that consumers will get “personal AI agents” and AI-powered smart glasses that can interact with the world in front of them.

These consumer-facing products are part of a broader strategy that uses large language models to rapidly build out Meta’s suite of social apps. Recent launches include an app for Marketplace sellers, another for Facebook Groups, one for vibe-coded games, and other experiments.

“I expect it to become a lot easier to ship new apps,” said Zuckerberg. “So we are planning to build out more ideas and use our recommendation systems to scale them to the people who will find them interesting.”

More experiments are on the way. The message is clear: Meta is betting big on AI across the board — enterprise, consumer, and everything in between.

Why your AI subscription bill is about to spike

If you’ve been leaning on ChatGPT or Claude for daily tasks, you’ve probably noticed the creeping costs. What started as a $20 monthly experiment can balloon into hundreds — even thousands — of dollars as your team scales usage. The pricing models from major AI providers are designed to grow with your dependence.

But there’s another path. Open-weight AI models let you run powerful language models on your own hardware, cutting out the middleman entirely. No monthly fees. No per-token charges. Just a one-time investment in hardware and a bit of technical setup.

What exactly are open-weight AI models?

Unlike closed models like GPT-4 or Gemini, open-weight models have their trained parameters — the “weights” — publicly released. Developers can download them, run them locally, and even fine-tune them on proprietary data. Think of it as owning the engine instead of renting a ride.

Popular examples include Meta’s Llama 3, Mistral’s Mixtral, and the Falcon series from the Technology Innovation Institute. These models range from 7 billion parameters (laptop-friendly) to 70 billion (requires serious GPU power).

The key advantage? Once you’ve downloaded the model, inference costs exactly zero dollars per query. No surprise bills at month-end.

What hardware do you actually need?

Here’s the honest truth: running a 70-billion-parameter model locally isn’t cheap on the hardware side. You’ll need a high-end GPU with at least 24 GB of VRAM — think NVIDIA RTX 4090 or A6000. That’s a $1,500 to $5,000 investment.

But here’s where the math gets interesting. If your business processes 500,000 API calls per month through a provider like OpenAI at roughly $0.01 per call, that’s $5,000 monthly. A $3,000 GPU pays for itself in under a month.

For smaller operations, 7-billion-parameter models run comfortably on a MacBook M2 with 16 GB RAM. No extra hardware needed.

• 7B models — MacBook M2/M3, any GPU with 8 GB VRAM
• 13B models — RTX 3090/4090, Apple M2 Ultra
• 70B models — Dual RTX 6000 Ada, or cloud GPU rental

You can also use cloud GPU providers like Lambda Labs or RunPod for occasional heavy lifting — paying by the hour instead of per token.

Software setup: simpler than you think

Five years ago, running a local LLM meant compiling obscure Python libraries. Today, tools like Ollama, LM Studio, and LocalAI offer one-click installs. Download the app, pick a model from their library, and start chatting.

For businesses with technical staff, vLLM and Text Generation Inference provide production-grade serving with OpenAI-compatible APIs. That means your existing code that calls OpenAI’s API can point to your local server with a single URL change.

Privacy is another massive win. Sensitive customer data, internal documents, or proprietary code never leaves your network. No data used for training other companies’ models.

What about fine-tuning?

Open-weight models can be fine-tuned on your specific business data using tools like LlamaFactory or Axolotl. A customer support bot trained on your actual support tickets will outperform a generic model every time — and you own the result.

How to pick the right model for your needs

Not every business needs a 70B behemoth. Start with a simple framework:

1. Task complexity — Simple Q&A? A 7B model works. Complex code generation or multi-step reasoning? Go 13B or higher.
2. Latency requirements — Real-time chat demands smaller models. Batch processing can handle larger ones.
3. Privacy needs — If you handle HIPAA, GDPR, or PCI data, local deployment isn’t optional.

The Hugging Face leaderboard is a solid starting point. Filter by parameter count, benchmark scores, and license terms. Many open-weight models use permissive licenses (Apache 2.0, MIT), but some — like Llama — have specific commercial terms you should read.

The hidden costs most people miss

Local deployment isn’t free. Electricity for a 450W GPU running 24/7 adds roughly $40–$60 monthly depending on your rates. Hardware maintenance, cooling, and IT staff time need factoring in.

But compare that to an API bill that scales linearly with usage. If your AI usage grows 10x next year, your local costs stay flat. Your API bill grows 10x.

For businesses with predictable or growing AI workloads, open-weight models offer a ceiling on costs. The more you use AI, the more you save.

First steps to get started today

Download Ollama on a laptop. Pull a 7B model like Llama 3.1 or Mistral. Test it against your most common AI tasks — summarization, drafting emails, answering product questions. Compare the output quality to your paid service.

Most teams find that 80% of their daily tasks are handled perfectly by a local 7B or 13B model. The remaining 20% — complex reasoning, creative writing — can still use the cloud API. That hybrid approach alone cuts costs by 60–80%.

Open-weight models aren’t a futuristic concept. They’re ready now, they’re getting better every month, and they could save your business thousands starting next billing cycle.