Perplexity’s DeepSeek Advantage: Why Your Monthly Bill Won’t Budge

Let’s pretend we’re sharing a corner table at your favourite coffee spot inside a bustling start-up incubator. Laptops, half-empty espresso cups, and ideas are flying around. Amid this caffeinated chaos, I drop a nugget of tech insight: Perplexity, that AI-powered search start-up making waves, has just made a big move with an open-source model called DeepSeek-R1. It’s the kind of move that lowers costs as dramatically as a craft cocktail bar sourcing premium ingredients at bulk warehouse prices — think top-shelf flavours for a fraction of the cost. But those savings won’t be used to lower your subscription costs anytime soon. Instead, they’ll fund an ambitious roadmap that would catapult Perplexity from a scrappy competitor into the AWS of AI search. It’s clever, and it helps you, just not in the wallet.

Let’s break it down.

Setting the Stage: Craft Cocktails Meet AI Search

Perplexity started out gaining a following for delivering Google-level immediate search results alongside advanced AI insights — summaries, data visualisations, “chain of thought” reasoning, and more. Traditionally, Perplexity leaned on premium, licensed large language models (LLMs) such as GPT-4 or Claude. These partnerships put them in the same bucket as many AI-based start-ups that rely heavily on big-name providers and pay fees per query (or token) processed.

But in AI, as in mixology, ingredients matter. You can pour a fancy $40-a-bottle top-shelf whisky or a $10 “house brand” that still tastes surprisingly smooth. Until recently, Perplexity was locked into paying the big brands every time they concocted an AI “drink” for their users. Those costs add up. For instance, renting OpenAI’s GPT-4 can cost around $15 per million tokens. At the scale of a popular platform, that’s tens (or hundreds) of thousands of dollars each month just to keep up with user demand.

Enter DeepSeek-R1 — an open-source large language model with custom architecture and an eye-popping reduction in operational overhead. With a name that sounds more like a futuristic spaceship than a whiskey label, DeepSeek-R1 promises to allow Perplexity to reduce licensing and compute costs significantly. It has, when used on its own, some very serious security flaws (mainly sending user data back to Beijing and censoring content). Those, according to Perplexity, no longer exist in their application of it. You could think of this Perplexity move as a cocktail bar discovering it can deliver the same top-shelf taste by distilling its own whisky in-house, eliminating multiple middlemen.

The Cost-Cutting Magic

Let’s get right to the core question: How exactly does DeepSeek-R1 save Perplexity so much money? We can break it down into three primary layers of cost avoidance, each acting like a different stage of a typical bar’s overhead:

1. The Licensing Tax Avoidance

Paying $15 per million tokens to OpenAI or other premium model providers is the AI equivalent of paying top dollar for branded spirits. Sure, you want that brand name to attract connoisseurs, but if you can offer a comparable in-house product, you might save a fortune.

By adopting DeepSeek-R1 as their house-brand whisky, Perplexity no longer shells out that token-based “rent.” This is the digital version of skipping the external liquor license fees. Suddenly, the overhead in mixing each “drink” (i.e., fulfilling each user query) is slashed dramatically.

Now, Perplexity can still offer “top-shelf” liquors (GPT-4 or Claude — you choose the model you want from a menu — for those with specialised needs — just like a bar might stock a few brand-name bottles for the true aficionados. But for the bulk of queries, especially from free-tier or standard subscription users, DeepSeek is good enough to keep them satisfied. This new approach parallels how streaming services often blend their own content with licenced blockbusters. They produce “house originals,” saving on licencing and distributing costs while still marketing a few big hits from external studios.

2. The Compute Diet

DeepSeek’s biggest technological advantage is something called Multi-head Latent Attention (MLA), which drastically reduces the “fuel” needed to run queries — in other words, the memory or compute overhead. It’s like putting a Prius engine in a Ferrari body: you still get the sleek, powerful chassis of a high-speed machine, but the energy efficiency skyrockets.

How big is the savings? According to early reference, MLA technology can reduce the model’s KV (Key-Value) cache memory usage by up to 93%. For the uninitiated, the KV cache is the model’s short-term memory, used for each token in a conversation. Reducing it is a massive step toward cutting the cost of inference (i.e., the cost to process a user’s request).

In practical terms, you can serve far more queries on the same hardware. If a standard LLM server can handle 100 queries a second before hitting memory limits, a DeepSeek-based server might handle 1,300 or more. That’s a big deal for any AI-based service. It means more “drinks” without hiring more bartenders or buying bigger bar countertops. I do think all the other providers will tailor DeepSeek R1 and implement it for this reason alone.

3. The Hidden Menu Effect

Have you ever visited a bar that offered a secret drink menu? Maybe you got a free taster of an experimental cocktail, which you loved so much you splurged on a full-priced version next time. In my pet topic of Behavioural Economics, that’s called the Hidden Menu Effect, a subtle psychological trick that’s been used for ages to drive upsells.

Perplexity’s free-tier approach does something similar. Users get three daily free queries — a small taste of what the system can do. But once they’re hooked on the convenience and insights, they might upgrade to a paid plan with 300+ daily searches. By having DeepSeek in its arsenal, Perplexity can afford to give away more freebies (or “samplers”) without burning a hole in its pocket.

DeepSeek thus becomes the gateway for building user loyalty. Since it’s cheaper and faster, those free queries don’t cost much. But once you’re sold on the brand, you might pony up for the advanced GPT-4 or Claude-based queries if you need deeper analysis, specialised reasoning, or just that extra “top-shelf” swirl.

Why Your $20 Bill Isn’t Shrinking

Now, right here is the original reason I wrote this article. All these cost savings — both from licensing and from compute overhead — sound like they should translate into an immediate dip in subscription costs, right? After all, if it’s cheaper for Perplexity to serve queries, they should pass the savings on to us, the users.

But again my the bar analogy continues to hold: just because the bar saves on wholesale liquor prices doesn’t necessarily mean your cocktail suddenly costs half as much. Instead, the bar might invest in ambiance, new bartenders, or marketing to attract more customers. Or they might just make more profit without doing any of those things. From what we can tell from the outside, Perplexity appears to be reinvesting these savings to grow, innovate, and position itself for a more expansive role in the AI marketplace.

Let’s look at three strategic reasons:

1. Portfolio Balancing

Despite the cost savings from DeepSeek, Perplexity still offers premium options like GPT-4 or the rumoured Claude 3.5 Sonnet (their poetic twist on Anthropic’s Claude). These advanced models remain expensive. It’s like a streaming service that wants to produce cheaper in-house shows and still pay licencing fees for big Hollywood titles.

But, by freeing up funds through DeepSeek, Perplexity can keep GPT-4 or Claude on the menu without having to raise prices further. So, for the same monthly subscription fee, you get that advanced AI when you really need it — while the cheaper model handles most day-to-day tasks. It’s a balancing act: keep the best of both worlds while staying competitive.

2. The Feature Arms Race

In AI-driven products, features are everything. Whether it’s advanced summarisation, chain-of-thought visualisations, voice interactions, or plug-ins, each new capability can help a platform stand out in a crowded market. Developing these features, however, isn’t free.

Perplexity has heavily touted visual aids for “chain-of-thought” reasoning, letting you see a breakdown of how the AI processes your query. This is an expensive feature to develop and maintain. These kinds of interactive UI components eat significant research and engineering resources.

But guess what’s footing the bill for these new goodies? DeepSeek’s cost savings. Rather than giving you a cheaper monthly bill, Perplexity invests the margin into new bells and whistles. That $20 might remain the same, but now you’ve got extra features that enhance your user experience — like a bar that starts adding fancy garnishes, new glassware, or a live jazz trio to your evening.

3. The Enterprise Gambit

The real money in AI often comes from enterprise customers — big companies that need specialised solutions and have the budgets to match. For an AI platform, having a cost-effective, scalable model is a massive advantage when pitching to corporate clients.

Reduced inference costs (via DeepSeek) make it feasible for Perplexity to offer enterprise clients robust API packages at a fraction of what competitors might charge. Some estimates suggest they could undercut enterprise token rates by up to 90%. That’s a game-changer when you’re trying to lure Fortune 500 companies that want custom AI search integrated into their workflows.

In short, if the bar can stock cheaper, in-house whisky, it can offer bigger event packages at a more competitive rate. Those big event bookings — like enterprise deals in the AI world — are where the real profits lie. Users at the bar might not see a discount on their individual drinks, but the overall volume and profitability go up, fuelling further expansion and improvements.

The Road Ahead: Where Perplexity Wants to Go

So, Perplexity is saving money on “ingredients” (licencing) and overhead (compute). Instead of passing that discount to end users, they’re likely using it to fund a multi-pronged push for market — and especially enterprise — dominance. Here’s Perplexity’s near-future blueprint as I see it:

1. Expanded Free Tier Capabilities

If you’re one of those casual users who only dips into Perplexity a few times a week, you might see your free tier get more generous. Maybe 10 daily queries instead of 3, or maybe a bigger monthly allotment. Why? Because now it’s cheaper to give away free samples.

And from a marketing perspective, free taste equals greater brand loyalty. If Perplexity can handle 10 times as many queries for the same cost, it’s no skin off their back to let you experiment with new features — so that eventually, you’re tempted to upgrade or at least recommend the service to others.

2. Vertical-Specific Models

DeepSeek’s architecture doesn’t just save cost; it also allows for more specialised training. It’s relatively straightforward to take an open-source LLM and fine-tune it for medical, legal, coding, or other specialised domains. That’s a big advantage over purely black-box solutions like GPT-4 or Claude, where you might have limited fine-tuning options or high additional costs.

Expect to see Perplexity roll out new “flavours” — a “DeepSeek-Med,” “DeepSeek-Legal,” and so on — without jacking up subscription prices. Each specialised model can attract niche audiences. Imagine a single AI subscription that includes a coding model, a medical Q&A model, and a general-purpose search model, all in one interface. That’s a powerful industry-specific lure for professionals and businesses alike.

3. Absorbing Future AI Regulation Costs

There’s a lot of talk about AI regulation on the horizon — data privacy, compliance standards, usage monitoring, and more. These regulatory frameworks could force AI companies to invest in compliance, data protection, or additional auditing. Many start-ups might pass these costs directly to consumers by raising subscription fees.

However, Perplexity’s cost savings via DeepSeek could give it a cushion to absorb some of these regulatory costs. That bar I keep talking about has enough margin from selling in-house whisky to afford new health inspections or licencing fees without hiking drink prices.

4. The Endgame: Becoming the AWS of AI Search

The ultimate ambition — if you look at Perplexity’s trajectory — is to become the AWS of AI search. AWS didn’t become the giant it is by giving you cheaper monthly bills on streaming services. Instead, it poured resources into building an ever-growing set of features, from simple web hosting to advanced machine learning platforms, which lock users into their ecosystem.

Perplexity could do the same with AI-driven search. As they save on operating costs, they’ll reinvest in more robust developer tools, plug-ins, API endpoints, and third-party integrations. Over time, they’ll become the “infrastructure layer” for AI search across multiple industries. That’s where the real growth (and real profits) reside. Well, this is my supposition.

What It Means for You

If you’re a casual user, don’t hold your breath waiting for your subscription to drop from $20 to $15. That’s not how this game is played. Instead, you’ll likely see:

• More free queries: Great for the hobbyist or the occasional user.
• Better features: Expect new ways to interact with your AI searches — visualisations, voice control, advanced citations, etc.
• Specialised verticals: A single subscription that covers your personal, professional, and niche queries.

For power users or businesses, the real draw might be advanced usage tiers, cheaper enterprise deals, or specialised models that deliver massive productivity gains.

And if you’re an AI enthusiast or competitor, or an AI Trainer like me, pay attention to the ripple effects. This is likely the beginning of a broader industry trend where open-source models eat into the market share of name-brand providers. We could see new variants of open-source LLMs that match or exceed proprietary ones in both quality and efficiency. That’s huge for driving innovation and lowering barriers to entry.

Final Thoughts: The Rising Value Ceiling

Will your $20 subscription feel any cheaper? Not literally. But in a value sense, you might be getting more for your money than ever before. Like a bar upgrading from peanuts to gourmet snacks, a live band, and a fancy garnish on every cocktail — your outlay remains the same, yet the experience is elevated.

Is DeepSeek via Perplexity safer than DeepSeek direct? Without a doubt. We do have to take Perplexity’s word for it that none of the data used in Perplexity, processed by DeepSeek, ends up in Beijing. But the proof of having hammered the censorship out of it is plain to see. Just using DeepSeek via Perplexity is vastly safer than using the DeepSeek app. I would never ever recommend using or installing the DeepSeek app directly.

Perplexity’s move with DeepSeek tells us a great deal about where the AI market is headed. It underscores how quickly open-source innovation can shake up an industry that was, until recently, dominated by a few big players. It also reminds us that in tech, cost savings don’t always trickle directly down to the consumer’s pocket. Instead, they often fuel expansions in quality, reach, and ecosystem lock-in.

So, if you’re sitting at the bar stool, gazing at the new house whisky on the shelf, remember: it might not be cheaper, but it’s probably a strategic masterpiece ensuring the bar (Perplexity) can keep pouring you new flavours, chasing new frontiers, and eventually positioning itself as a top contender in the AI world. Cheers to that.