The European Court of Justice finally closed a decade-old case this week: Apple must pay €13 billion in back taxes to Ireland. The decision reverses a 2020 ruling and confirms what was always obvious—that the company engineered its tax liability down to 0.005% by the early 2010s. The mechanism was elegant and legal: two Irish shell companies, one managing sales across the globe, another holding intellectual property. Every iPhone sold outside the US generated revenue in the sales company, which then paid licensing fees to the IP company, zeroing out taxable profit in Dublin before moving the money to a tax haven. By 2014, Apple paid virtually nothing on billions in revenue.
This Double Irish arrangement wasn't unique. Facebook, Google, Amazon all used variants. Ireland enabled it by allowing companies registered in Dublin but headquartered elsewhere to move profits tax-free. The scheme ended in 2015-16 under EU pressure, but the damage was structural. Ireland's 2015 GDP growth hit 26%—not because the economy actually grew, but because multinationals consolidated European operations there. The real picture emerges in GNI (Gross National Income), which strips out foreign-owned production: using that metric, Ireland ranks 12th in Europe, not first. Yet the political economy worked. Those 14 of Ireland's top 20 companies are multinationals; they employ a fifth of private-sector workers and pay 80% of corporate tax revenue. The cost: housing inflation, wage pressure, and an economy hostage to tech giant accounting decisions.
The €13 billion windfall creates a different problem. Ireland's budget was already €8.6 billion in surplus. Now the state faces pressure to spend money it doesn't need—a luxury problem, but real. The government will likely stash reserves for pensions and future crises, recalling how Covid drained the coffers in 2020. Meanwhile, this case signals the end of the race to zero. The EU's pressure worked. But it also shows how sophisticated tax engineering persists: entire departments across multinationals exist solely to find the next legal edge. The Leprechaun Economics era—named by Nobel laureate Paul Krugman—is closing. What replaces it remains unclear.
On AI, OpenAI's new O1 models signal a shift away from pure scale. Instead of bigger datasets and more parameters, O1 introduces chain-of-thought reasoning: the model thinks through problems step-by-step, showing its work before answering. On Olympic-level math problems, it scores 83% versus GPT-4's 13%. The "O" doesn't stand for GPT-5; it marks a reset in how the company names things. This matters because it suggests OpenAI believes the next frontier isn't more training data—it's different training methods. The model likely uses something like adversarial training, where two models challenge each other until convergence, similar to how diffusion models work in image generation. The preview is available only to ChatGPT+ users and high-tier API customers; the full rollout and the lighter O1 Mini remain months away. But the framing is telling: this is positioned as a new era, not an incremental update. Whether that proves true depends on real-world utility, which remains to be seen. For routine queries, GPT-4o stays faster. For scientific reasoning—genetics, quantum physics, economics—O1 could accelerate hypothesis generation. The implication: AGI won't arrive through scale alone. It will come through rethinking how models learn.