Nvidia hit a $3 trillion market cap in June 2024. For a few days it was worth more than Apple. A chipmaker, however dominant, briefly outvalued the company that defined consumer computing for two decades. The forward price-to-earnings ratio sat north of 30x, nearly double the S&P 500 average. South African fund managers holding Nasdaq trackers through local ETFs saw this reflected in their statements without necessarily grasping what they were betting on: they were betting that AI infrastructure spending keeps accelerating forever.
That bet has started to look familiar.
The Pattern We Keep Repeating
Every technology cycle produces the same choreography. A genuine breakthrough emerges. Capital floods in. Valuations detach from near-term revenue. Late entrants pile in on FOMO. Then someone asks where the profits are, and the music stutters. The dot-com bubble had Pets.com and eyeball metrics. The crypto boom had white papers and no users. AI has actual working products, buried under layers of speculative excess.
OpenAI reached an $80 billion valuation in early 2024 through a private tender offer. Anthropic, its closest rival, collected over $7 billion from Amazon and Google to reach $18.4 billion. Microsoft’s total commitment to OpenAI reportedly exceeds $13 billion. These are not venture capital seed cheques. They are strategic bets by the most cash-rich companies on Earth, made at prices that assume near-total market capture. The technology works. ChatGPT writes code. Claude summarises contracts. The question is whether any of this generates returns commensurate with the capital deployed.
The South African angle is indirect but real. Local businesses buying Microsoft 365 Copilot licences, or paying premium rates for AI-enhanced SEO tools, are funding this arms race through their subscription fees. They are also absorbing the hype through their marketing budgets, often without clear metrics for what changes.
What AI-Washing Looks Like on the Ground
Walk through any SaaS marketplace now and every second product has an “AI-powered” badge. Some of these are genuine integrations: a content tool that actually calls GPT-4 to generate drafts, a support platform that routes tickets through a classification model. Others are spreadsheet formulas with marketing budget. The pattern is old. In 2023 it was “blockchain-enabled.” In 2021 it was “Web3 native.” The label attaches to anything that needs to justify a price hike or attract investor interest.
For practitioners, the cost is concrete. A Johannesburg-based digital agency recently told me they spent R180,000 on an “AI content platform” that turned out to be a thin wrapper around OpenAI’s API, marked up 400%. The vendor had no proprietary technology, no fine-tuning on local language patterns, no integration with their WordPress stack. They could have built the same workflow with Make.com and a $20 ChatGPT subscription. The sales deck had used the right words, and the competitive pressure to “do something with AI” was intense.
This is the practical risk of an overheated market: not that AI fails, but that businesses overpay for mediocrity while the genuine article sits one API call away. Value gets destroyed in the gap between transformative technology and transformative investment returns.
The Infrastructure Bet vs. The Application Bet
Nvidia’s valuation surge rests on a real physical constraint. Its H100 GPUs cost $30,000 to $40,000 each, and demand outstrips supply. Data centre operators from Cape Town to Frankfurt are scrambling for capacity. This is unlike the dot-com bubble in one crucial respect: there is tangible hardware, with tangible scarcity, generating tangible revenue. The risk is not that the infrastructure disappears. It is that too much gets built too fast, and the returns compress.
The application layer is shakier. Foundational model companies burn cash at rates that make Uber’s early years look conservative. Training runs cost millions per attempt. Talent salaries have inflated beyond what any normal business can match. The bet is that one of these companies achieves sufficient scale to become the default platform, the next AWS or Android. But platforms are rare, and the current field has half a dozen credible contenders plus the tech giants building internally. Not all will survive. Not all need to for the technology to succeed. Many investors will lose money regardless.
South African businesses are mostly application-layer buyers, not infrastructure investors. They rent compute, subscribe to APIs, buy tools. Their risk is not that Nvidia collapses. It is that they build workflows around a model or vendor that pivots, gets acquired, or prices them out once the subsidy period ends.
A Practical Evaluation Framework
The antidote to hype is not scepticism for its own sake. It is specific questions, asked consistently. Here is a checklist for evaluating any AI tool or vendor before budget commits:
Define the problem before the solution. What exact workflow slows down? What metric moves if this works? “Improve content” fails. “Reduce first-draft production time from six hours to ninety minutes, measured across ten articles” works. Vendors who cannot engage with your specific metric are selling hope.
Demand vendor financial transparency. How old is the company? What is their funding runway? Who are their other customers? A twelve-month-old startup with one seed round and no named references may deliver brilliant technology, but they may also disappear or get acqui-hired into oblivion. For core workflows, prefer vendors with enterprise support infrastructure and clear SLAs.
Audit the data flow. Where does your input go? Does it train the vendor’s models? What jurisdictions govern their servers? For South African businesses handling POPIA-regulated data, this is not optional. Many AI tools default to processing in US or EU regions with unclear compliance pathways. Ask directly. Document the answer.
Test for model agnosticism. Can you swap the underlying LLM without rebuilding your workflow? Tools locked to a single provider expose you to price shocks and capability gaps. The API-first architecture, where your application calls whichever model is currently best for the task, provides flexibility that proprietary platforms resist.
Measure for thirty days, decide for ninety. Pilot with real work, not demo data. Track time saved, error rates, and any new friction introduced. AI tools often speed one step while complicating three others. The net gain is what matters, not the headline feature.
Building Internal Capability
The most resilient AI strategy is not vendor selection. It is internal literacy. A team that understands how prompting works, how models differ, and how to evaluate output quality can adapt as the market shifts. A team that delegates all of this to a platform becomes hostage to that platform’s roadmap.
This does not mean hiring machine learning PhDs. It means investing in practical education: prompt engineering for content teams, API basics for developers, bias recognition for decision-makers. The cost is modest against a single failed vendor contract. The compound return is a business that can distinguish signal from noise as the cycle evolves.
South Africa’s technical talent pool is deeper than often assumed. Local universities produce capable computer scientists. The constraint is usually organisational: businesses that view AI as a purchased solution rather than a learned capability. The former approach works until the vendor changes terms. The latter approach compounds.
What Happens When the Cycle Turns
Market corrections in technology do not typically kill the underlying innovation. The dot-com crash destroyed Pets.com but left Amazon and Google. The crypto winter eliminated thousands of tokens but preserved the blockchain concept for specific use cases. An AI correction, if it comes, would likely consolidate power among the best-capitalised players while clearing out the wrapper companies and speculative valuations.
For practitioners, this is not a reason to delay adoption. It is a reason to adopt carefully. The businesses that thrived after previous corrections were those that used the technology productively during the boom, without betting the company on any single vendor’s success. They treated AI as a capability layer, not a strategic identity.
The current investment boom will leave behind genuine advances: better models, cheaper inference, more reliable tooling. It will also leave behind burned budgets, abandoned platforms, and businesses that paid premium prices for capabilities that became commoditised within eighteen months. The difference between those outcomes is not luck. It is the discipline to separate what works from what is merely well-funded.
